00:00:00 - 00:14:30 - Breakout Session: Foundation models introduction and presentation (Fabian Theis, Bo Wang)
00:14:30 - 01:16:06 - Group discussion
Session summary:
This session on foundational models (FMs) highlighted their tenacity for data representation and analysis along the central dogma of molecular biology. The session highlighted several advantages to FMs, including that they are effective at transfer learning, their generalized framework allows models to competitively work on multiple tasks, and they provide a compatible architecture for data integration and adhere to scaling laws. However, they also suffer from lack of interpretability, high compute power, flawed associations and model accessibility challenges. The breakout session attendees agreed on the critical need to include biologists in model development and the need to gather human-machine interaction metadata to construct robust FMs. This intrinsic human element also means that the ethics of FMs must be safe-guarded in data mining and curation, legal and copyright issues and algorithm hallucination.
![](https://i.ytimg.com/vi/B4E6qmnuu1M/mqdefault.jpg)