Bogdan Didenko Introduces an Open Atlas of Biological Data Routes into Generative Models
Bogdan Didenko Introduces an Open Atlas of Biological Data Routes into Generative Models
On August 23, Bogdan Didenko introduced an atlas that shows the forms in which biological data reach generative models. The current snapshot contains 55 records, 109 models, 467 task configurations, and 585 documented input routes. Generative models produce new text, images, or other data. The atlas helps identify the signal that such a system receives.
This spring, a map of 280 biomedical foundation models organized the field by data type and the connections among those types. Didenko’s atlas examines the next step: how a specific task transformed DNA, RNA, or text before passing it to a model.
An RNA profile, which is a set of measurements of gene activity in a cell, can be passed to a model as a list of genes, a sequence of symbols that encodes biological information, or a numerical vector, meaning an ordered series of numbers. In each case, the model receives a different input. The atlas traces the full route: which biological object was used, how it was transformed, and the form in which the model received it.
A single study may use one model for several tasks, with each task receiving its own input. The atlas therefore counts task configurations, meaning specific tasks paired with specific input forms. Multiple format families appear in 40 of the 55 records, but they are combined within a single configuration in only 46 of the 467 cases.
The label “multimodal” describes a study as a whole. The atlas shows the input form used for a specific task. It therefore allows readers to check whether a system combines several data representations or works with only one.
Each route includes an excerpt from the original study, its title, and the pages that describe the model input. The project repository contains the review protocol, corpus, scripts, and a fixed snapshot of the data. These materials allow readers to trace a category on the map back to the publication on which it is based.
A claim that a model is multimodal can now be examined through three questions: what signal entered the system, how it was transformed, and the form in which the model received it.