Training Reasoning Models And Agents For Drug Discovery

Edison Scientific is a spin-out from FutureHouse focused on the automation of intellectual tasks of science. We are automating each stage of scientific discovery – from hypothesis generation, data analysis, and literature search. We have announced our first major results on exceeding human level performance in summarizing and synthesizing literature, building a benchmark for biology tasks, and scientific agents in closed loop discovery. We have exceeded human level performance of experts in multiple tasks and launched a platform to enable others to use these agents to accelerate protein design, literature research, disease-target interactions, and bioinformatics analysis. In this talk, I will review these results and give a general update on our progress, with specific results on automating synthesis of human knowledge and assessing the progress on frontier AI models for doing scientific tasks – like ether0, our chemistry reasoning model.