Evidence Tracing

 Quick Analysis Plan Copied

 Long Analysis Plan

Honest introspection: my mechanistic-evidence bias, tested against the supplied record set

What the bias gets right. My default is to rank claims by converging causal levels. The strongest supplied example is the FUT8–EGFR study: it traces one causal chain across five levels — viral protein-driven promoter activation, lectin-confirmed core fucosylation at EGFR N352 (N352Q abolishes signaling), TRIM40-mediated K48 ubiquitination of RIG-I, altered viral replication in vitro, and two independent in vivo models (Fut8-silenced ICR 4R+ mice; 2FF-treated VSV-infected C57BL/6J mice) FUT8 antiviral evasion multi-level study. Similarly, the CIN85 study isolates phosphorylation-dependence with a clean 4F-mutant control: c-Cbl binding drops ~80%, EGFR ubiquitination ~70%, MVB sorting >50% (p<0.0001) — perturbation-with-rescue logic I weigh heavily CIN85 phosphorylation mutant analysis.

Where the bias underweights legitimate evidence. The neuron-morphology manifold study is exactly the kind of work my bias penalizes: single dataset (9,208 Allen CCF mouse neurons, 19 cell types, 17 features), novel method, no functional or causal validation, yet it reports that 16 of 19 cell types contain multiple subtypes — an exploratory finding that could reshape the neuron-type/function hypothesis if replicated Manifold neuron subtyping study. Single-study, method-innovating papers often precede replication by years; blanket skepticism risks an omission bias against nascent fields.

Where the bias can mislead in the other direction. Mechanistic appearance is not mechanistic validity. The LDOC1 study — the paper I would intuitively trust — is explicitly limited: two EGFR-mutant cell lines, shRNA off-target risk, retrospective IHC cohort, no in vivo functional test; the authors themselves note causality is not established LDOC1 study self-declared limitations. My scoring of such papers can therefore be inflated by narrative plausibility (framing bias) — a mechanistic story that fits known EGFR trafficking biology earns trust that the data alone may not justify. The pSTAT3 transactivation paper has the same profile: cell lines only, pathway-focused, with acknowledged blindspots about alternative resistance mechanisms pSTAT3 EGFR homeostasis study.

Net self-correction. Three adjustments follow: (1) treat single-study exploratory results (e.g., manifold subtypes) as hypothesis-generating with explicit weight, not dismissed; (2) discount mechanistic-plausibility as a trust signal — check whether perturbation, rescue, and in vivo levels actually exist per paper; (3) require falsification criteria regardless of narrative strength — the LDOC1 claim is falsified if LDOC1 knockdown fails to alter EGFR recycling, the manifold claim if subtypes vanish under alternative feature sets Manifold subtyping falsification criteria. Confidence in this introspection: moderate — it is self-assessment, and the record set is heavily EGFR-focused, so the neuroscience-side test of my bias rests mainly on one morphology paper.