About HiveBlot

HiveBlot turns published Western blot figures, captions and methods into structured, searchable, auditable experiment-level evidence: which protein, modified where, in which cells, under what treatment, detected with which antibody, and what the blot showed lane by lane.

Who it is for

Wet-lab researchers who need to locate and compare experiment-level Western blot evidence. The sharpest questions it is built for: has this antibody or catalog number already appeared in published work, and what did the blot look like? Has this protein been blotted in this cell line under this treatment? Can I trust this row, and where exactly did each value come from?

What the beta contains today

...
REVIEWED PAPERS
...
EXPERIMENT GROUPS
...
EVIDENCE ROWS

Loading corpus statistics...

The corpus is intentionally small and reviewed. Every row was checked field-by-field by a human and then by an independent scientific QA pass. One row is one lane of one blot panel; rows group into experiments. Selected papers from the reviewed corpus:

A search that returns nothing usually means the topic is not in the reviewed papers. It does not mean no such evidence exists in the literature, and the search page says so.

What an Evidence Record is

Not a single answer, but a container. Each important field (protein identity, phosphosite, cell line, organism, treatment, dose, duration, antibody, molecular weight) carries its own value, a status, the source it came from (caption, methods, antibody entry, the model's own reading, UniProt) and the wording that supports it. The normalized value and the wording as printed in the paper are both kept, so every normalization is auditable and reversible.

What HiveBlot refuses to do

  • It does not guess. When credible sources disagree, the field is marked Conflicting, keeps no value, and shows both claims.
  • It does not fill gaps. A value the paper never states is shown as not reported, never as a plausible number.
  • It does not measure intensity. Band presence is categorical (present / absent / uncertain). Evidence Records carry no densitometry, and multiplicity such as a doublet is recorded descriptively only. (An uncalibrated band reading page exists under Experimental; it measures within one image, refuses clipped or unstable readings, and stores nothing.)
  • It does not read mass off an image. Expected molecular weight comes from the UniProt sequence; reported weight comes only from what the paper printed. They are stored separately.
  • It does not overwrite itself. Researcher corrections are stored beside the extraction as an auditable pair, never applied over it.

What this beta proves, and what it does not

Demonstrated on the reviewed corpus

  • Structured Evidence Records with field-level provenance
  • Experiment-level search (protein, phosphosite, antibody, cell line, co-IP, loading control)
  • Explicit uncertainty and abstention
  • Conflicts preserved rather than silently resolved
  • Researcher feedback stored beside model output and surviving reloads

Not demonstrated yet

  • Coverage of the wider literature beyond the reviewed corpus
  • Automated extraction accuracy on papers HiveBlot has not been shown; the model-reading step for the selected papers above ran with a human in the loop
  • Precision, recall or F1 on unseen papers: no such number exists yet
  • Densitometry or any quantitative band-intensity inference

The reviewed corpus is the baseline that automated extraction will be measured against. Until that evaluation runs, no accuracy claim about new papers is made anywhere on this site.

Beta hosting note

The beta runs on free-tier hosting. The first search after a quiet spell can take 30–60 seconds while the API wakes up; everything after that is fast. A slow first search is the hosting plan, not the product.