I run a research program at BiodynAI aimed at advancing mechanistic interpetability of bioligical foundation models. I mean models trained on things such as DNA sequences, proteins, gene-expression profiles, cell images, tissue samples, spatial measurements, and other biological data. Many of them learn through some version of “hide part of the data and predict it back”, although the exact objective varies.

There are 3 cases for that:

1. Bio AIs are "model organisms for mechinterp in the wild": they provide rather clean and clear ground truth, they are relatively small, and yet they are real models and real useful things can be done with them, including things that generalize to LLMs.
2. To do biosecurity audit of bio AIs, we desperately need mechinterp. These models learn a lot of knowledge which is not accessible or derivable from their default outputs simply by design of those outpus. They may output sequences or RNA expression levels, but to predict those, they learn a lot of biology! That biology can be dangerous. The models can be fine-tuned using this knowledge. To understand the real capabilities of the models, we need mechinterp.
3. We desperately need to amplify natural human intelligence to help with alignment. My take is that we can use biological AIs and their learned internal representations to discover novel biology on scale and then use it to develop intelligence-amplifying interventions.

I think I have scientific evidence backing all of these claims and doubling down on this research agenda. Even if some of the claims above are false, I think it is still very promising and highly underappreciated research area.

I need courageous people to help with that. As the research gains traction, sometimes people come to me and ask what they can do. Initially, I proceeded on case-by-case basis, but now I am at the point where it is worth having some centralized list of open research questions to work on. Therefore, I created it.

So, meet the list of 109 open problems in biological mechanistic intepretability: https://biodynai.com/109-open-problems-biological-mechanistic-interpretability.html

This is meant to be updated regularly so I would keep it at the organization website.

If you start working on any of these problems, please get in touch with me so that we can coordinate. And of course, if you want to join our research group, you are most welcome. We don't have funding right now, so I won't be able to pay yet, but I am looking for funding.

I believe we achieved quite cool results, and it is worth scaling them. In particular, we:

• Achieved best performance on gene regulatory network inference from single cell models using mechinterp and overall.
• Trained so far biggest tensor networks on single cell data and demonstrated that they are both more intepretable and performant.
• Generated a bunch of hypotheses for wet lab validation.
• Did a lot of mechinterp evalutations of frontier biological models.
• Automated substantial part of research with agents relying on validation on biological ground truth.