)
)
How Can We Make LLMs Even Better at Drug Discovery?
LLM are very versatile models and offer a natural multimodal interface between drug scientists and the variety of datasets they need to interact with. We explore different key aspects for Drug Discovery where LLMs can be improved, namely their ability to retrieve and reason on structured data (tables and graphs) and their ability to act as predictive models on such data.
In particular, we will cover the following problems in the context of target discovery:
- Gene perturbation predictions by finetuning LLMs on biological data
- Zero-shot patient outcome prediction from EHR tables.
- Biological reasoning on human data tables and graphs
