http://www.unidata.ucar.edu/software/netcdf/ http://cdf.gsfc.nasa.gov/
http://download.cell.com/images/journalimages/2211-1247/PIIS... http://download.cell.com/images/journalimages/2211-1247/PIIS... http://download.cell.com/images/journalimages/2211-1247/PIIS...
Most science (that really needs hackers' help) is not the big sexy datasets, but the really basic HTML5/CSS viewable graphs otherwise done in Excel. But it must be done in a scientifically rigorous way (retain units and labels, be able to manipulate those units, and be able to show/hide data between different datasets).
Think a way to download and view a melting point curve dataset without ANY user configuration. Or an enzyme activity curve. Or a drug dosage response curve. VERY basic data, but it must be able to maintain and interpret a label and a unit in addition to the number itself.
In the examples in your final paragraph, surely what is needed there is a new tool, not a new data format?
I think the larger 'problem' (if it really is a problem, and not just hackers wanting to believe they know the One True Way) is that there is no such thing as the typical 'scientist'. Every small sub-discipline has its own needs, its own standards that it has settled on and its own toolchain. There is definitely room for improvement within every one of these small areas, but there's no global problem that needs solving.