I’m most interested in supervising technical research projects that overlap with our broader research program at Fourth Eon, where we are evaluating biological AI models for function-based biosecurity screening. Examples include the following (open to scholar suggestions):
- Applying mechanistic interpretability techniques to protein foundation models to probe their learned representations of functionally meaningful biophysical and biochemical properties (e.g. local structure, surface charge, catalytic activity, solvation). The goal is to understand how semantic information is encoded in specific features and circuits for different models, and how to leverage them for generalized hazardous function prediction.
- Evaluating bio model capabilities for predicting function of non-natural genes and proteins that may be outside the training distribution, using a curated test set of engineered sequences tied to experimental validation data. This provides the tools to determine model utility for detecting engineered or AI-designed threats using function-based screening.
- Developing multi-modal sequence analysis workflows for use in adaptive biosecurity screening. This project explores the integration of different model capabilities (e.g. structure prediction, binding, function, toxicity) and synthesize the outputs to produce biosecurity risk assessments on unknown sequences.