I've been using GTD in biomedical science and academics for about a year. Though it's not particularly science-specific, some might be interested in a description of my approach, using Mori, at http://jhh.med.virginia.edu/main/MoriGTD.

There are two issues in GTD that current software isn't addressing well and that are relevant for implementations of GTD in science and academics:

1. If you implement GTD with any sort of comprehensiveness, you will have a large number of projects--as David Allen points out in his book. I have more than 30 major projects and that's because I've slipped a bit recently. Over 50 wouldn't be unreasonable. It's impossible to review all of these projects in detail every day, and, in fact, many don't need it--but some do (a wide variety of different types of projects with differing requirements is common in academics). Projects need varying levels of attention, depending on their nature and where they are in their own lifecycles. Software needs to address this, with the ability to set project review schedules that slip the projects into your daily list as needed. These schedules should also be easily modifiable during project review so that review frequency can track actual project activity. Software should also catch projects that are being overlooked and fail to show activity within a set period of time, sort of like how NetNewsWire catches RSS feeds that are not being modified.

2. GTD apps should be a nerve center for projects, but should not necessarily try to store all project-related documents internally. In addition to "multidimensional to-do lists" GTD apps should also manage dynamic data about projects (notes), links to documents and annotations of documents. There are a large number of document types in science and academics, including text of various types, several forms of markup, word processing, PDF, spreadsheets, etc. In my experience it's not particularly useful to pull all of those document types into a database in a GTD app. Rather, they should be managed externally with links to locations in the file system or the web and excellent features for link management. As above, locally-stored data should include the most dynamic data about projects (notes, to-tos and annotations of linked documents), links and possibly indexes of the content of linked documents.