Interview
Arc Institute's Patrick Hsu on Building an App Store for Biology with AI
- Machine learning in biology is expected to evolve from drug design toward a unifying theory spanning scales from planets to molecules, with specific milestones including the designable IgG antibodies with point-and-click enzyme precision by end-of-2025 and significant maturation in de novo enzyme design over the next couple of years.
- ARC's EVO model series aims to connect biological sequences to function, enabling the design of new CRISPR systems and improved prediction of coding and non-coding mutations, while the ARC Institute intends to build a virtual cell atlas as the world's largest single-cell dataset.
- The industry probability of success for drugs is projected to rise from 10% to 20%, 30%, or 50% if predictive models replace guess-and-check approaches, with immediate improvements anticipated in discrete pipeline steps like target ID and literature review.
- Predictive capabilities are expected to scale to multi-parallelized simulations of up to 10,000 agents acting as a trusted oracle, with a goal of 99.9% simulation accuracy for drug-target impacts validated in automated wet labs within hours by 2050.
- By 2030, accurate and useful virtual cell models are expected to exist that evoke a reaction from biologists, potentially integrating multimodal data for personalized "AI doctor" recommendations regarding diet, sleep, and behavior.
- Starting in 2025, AI agents will focus on the meta-aspects of scientific operations, deploying as co-pilots from hypothesis generation to data analysis, though fully automating the loop to write papers remains distant.
- ARC prioritizes tangible product capabilities and real-world disease cures over academic publications, leveraging a "mothership" structure of blended academic and industrial staff to attack long-term research breakthroughs.
- An "app store" ecosystem is expected to emerge around EVO's foundational layers, allowing individuals at pharma companies or new entrants to utilize the open-source model, while deep research models remain the most useful AI applications for immediate daily work.
- Researchers will continue focusing on multi-system interactions at the neuro and immune interface and physiological programming, with future human experimentation capabilities requiring regulatory innovation beyond mere data collection.