Biomolecular foundation models, such as AlphaFold3, Boltz2, and Chai-1, are emerging as powerful tools for structure-based drug discovery, capable of predicting drug-target interactions with high accuracy, and in some instances, the binding affinity. Ion channels are canonically difficult yet attractive targets for pathologies such as epilepsy, pain, itch, and cardiac arrhythmia. Current drug discovery strategies aim to selectively modulate channel activity by targeting specific domains of the channel, which are broadly categorized as the selectivity filter, the central pore cavity, and the voltage-sensing domains. To our knowledge, biomolecular foundation models have not yet been systematically evaluated on ion channel-ligand interactions. Here, we curated a data set of experimentally resolved ion channel-small molecule structures and assessed model predictive confidence, small molecule pose overlap with structure coordinates, and binding affinity when available. Our results demonstrate that biomolecular foundation models accurately recapitulate the interface at well-defined pockets, such as the voltage-sensing domain, but struggle to represent larger, dynamic surfaces, such as the central pore cavity. These findings highlight both the promise and limitations of biomolecular foundation models for ion channel drug discovery, underscoring the importance of benchmarking each model to ensure its suitability for confident prediction on the target of interest.
Harris et al. (Sun,) studied this question.