Aaron Maiwald

This paper evaluates the biosecurity risks of generative genome design models by quantifying evolutionary novelty and design efficiency in bacteriophage sequences generated by Evo 2.

This paper introduces Decode-gLM, a toolkit using sparse autoencoders to interpret, audit, and steer genomic language models, revealing biological features and identifying data contamination.

This study examines how scoring context and training data influence protein design models' preferences across species, showing that backbone-conditioned models favor archaeal and thermophilic proteins while sequence-only models favor eukaryotic proteins.