Arc Institute's Patrick Hsu on Building an App Store for Biology with AI

Arc Institute co-founder Patrick Hsu explains how Evo 2, a DNA-native foundation model, reaches state-of-the-art on variant pathogenicity and enables an app-store ecosystem for biology.

• Evo 2 is an autoregressive long-context model trained on trillions of genomic tokens spanning bacteria, viruses, fungi, and humans — no lab-in-the-loop RL used.
• Model is state-of-the-art at classifying BRCA1 variants of unknown significance, directly informing decisions like prophylactic mastectomy.
• Arc released the world’s largest single-cell atlas, built by two researchers plus one cloud crawler agent reprocessing all public single-cell data.
• Drug industry probability of success is ~10%; Hsu argues even moderate AI predictive power would be transformative by enabling massive parallelization instead of linear animal studies.
• Greg Brockman joined Arc during his first-ever OpenAI sabbatical and immediately asked to be added to the repo mid-meeting.
• Arc’s CTO Dave Burke previously ran engineering for Android and Pixel; Arc deliberately hires industry operators alongside academics.
• 2025 prediction: one-shot design of full IgG antibodies with correct CDR regions; 2030 prediction: virtual cell models that make cell biologists feel what NLP researchers felt post-GPT-3.

2025-04-15 · Watch on YouTube