BiodynAI lists 109 open problems in biological AI interpretability

BiodynAI, led by a researcher focused on mechanistic interpretability (mechinterp) of biological foundation models, has published a list of 109 open problems to advance transparency in models trained on DNA sequences, proteins, gene-expression data, and more. These models, which learn internal representations of biology through objectives like masked data prediction, often act as 'black boxes' despite their utility in tasks like gene regulatory network inference.

The research argues that mechanistic interpretability is critical for biosecurity audits, as these models may learn dangerous biological knowledge not visible in their outputs. By decoding these internal representations, BiodynAI claims to have achieved state-of-the-art performance in single-cell gene network inference, trained the largest tensor networks on single-cell data, and automated research workflows using AI agents. The team is actively seeking collaborators and funding to scale these efforts, positioning biological AI interpretability as a bridge to amplify human intelligence and improve AI alignment.