Key points are not available for this paper at this time.
The excitement around and usage of artificial intelligence (AI) tools in scientific research is increasing across fields, but lax publication standards are resulting in papers "like grand mansions of straw, rather than sturdy houses of brick" (Kaelin Jr). In biomedicine, AI methods span fundamental biology to clinical applications, such as prioritizing chemicals as drug candidates, deciphering transcription factor-DNA binding preferences, detecting RNA modifications, learning the effects of cell treatments, and linking medical images with text annotations. Despite repeated insistence from experts that sharing AI training data, code, and model weights-the saved parameters required to make new predictions-are necessary best practices, they are not universal practices among authors or publishers. A mere 10% of surveyed biology journals require sharing analysis code. This abysmal status quo needs to be changed. Experimental biology communities show there is a path forward. Depositing macromolecular structure data in the Protein Data Bank was initially voluntary. Leadership from structural biologists about the importance of mandatory data deposition led to the creation of policies and their subsequent uptake at journals. Likewise, genomics researchers pioneered data sharing principles that guided funding agency policies. The scientific community must show similar resolve in establishing and uniformly applying policies for open AI in biomedicine.
Gitter et al. (Mon,) studied this question.