Research
Current genome model is an attention-based model pretrained on billions of genes, with a classifier bolted on top. We think that biology models should be physics-informed neural networks, grounded in the truth of how biology actually works, capturing the complex dynamics of interactions.
Current genome foundation models may not be sufficient for function, phenotype, or the behaviour of an organism in its environment. They plateau with scale and often fail to beat simple baselines. Labs across world models, self-supervised learning, robotics, and multimodal AI are building architectures around persistent internal state and interaction with the environment. Genome models should be no different, built to capture a living cell’s complex, dynamic interactions.