Apr 14, 2025, 8:00 PM ETInterview

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

Core Technology: Patrick Hsu and the ARC Institute developed EVO2, a generative AI foundation model trained on genomic sequences (DNA) across all domains of life, rather than focusing solely on proteins.
Mechanism: The model utilizes an autoregressive, long-context architecture that predicts the "next base" in a sequence, allowing it to learn higher-order patterns of molecular logic and biological function directly from the genetic code.
Primary Advantage: By training on the fundamental information layer of life, EVO can reason over coding and non-coding mutations to predict their functional consequences, addressing the field's most persistent challenge.
Key Capability: EVO achieves state-of-the-art performance in interpreting "Variants of Unknown Significance" (VUS), distinguishing between benign mutations and those causing disease in genes like BRCA1.

Problem Solved: Current genetic testing (e.g., 23andMe) frequently returns VUS, leaving clinicians without actionable data for millions of patients.
Model Insight: EVO provides probabilistic assessments for VUS, effectively predicting whether a specific mutation is pathogenic or benign.
Verification Method: The team utilized ClinVar, a gold-standard database of known disease-causing mutations, to rigorously evaluate the model's predictive accuracy against ground truth.
Impact: The ability to reclassify VUS could prevent unnecessary medical interventions (e.g., double mastectomies) for patients with benign variants while ensuring high-risk patients receive early intervention.
Open Source Status: The EVO model is open source, aiming to catalyze an ecosystem of "apps" and tools for biology rather than creating a proprietary walled garden.

Philosophy Shift: Hsu argues that machine learning in biology should not be limited to drug design but must aim for a "unifying theory" that explains evolution and biology across all length scales (from ecosystems to molecules).
Data Strategy: Unlike protein models that rely on lab-generated data, EVO leverages the "experiment of evolution" by training on 25+ years of publicly available genomic data from the Sequence Read Archive.
No Lab-in-the-Loop: The model was trained purely on existing sequence data without Reinforcement Learning (RL) from wet lab experiments, relying on the predictive power of natural variation.
Interpretability Challenge: Because "DNA speaks" differently than human language, the team is prioritizing the development of interpretability tools and annotators to understand model reasoning, similar to how NLP handles grammar.

Mission: ARC (Amplify Research, Cure) is designed as a multidisciplinary "mothership" to attack long-term research breakthroughs by blending academic rigor with industry product sensibilities.
Talent Strategy: The institute recruits "bilingual" scientists who bridge the gap between biology and AI, noting that true translation capability between disciplines is rare.
Incentive Structure: Unlike universities that optimize for publications, ARC focuses on tangible outcomes, technical blogs, code repositories, and technologies with real-world product potential.
Collaboration Model: The organization integrates faculty from Stanford, Berkeley, and UCSF with industry experts (e.g., former OpenAI staff) to foster cross-pollination of ideas and operational efficiency.

Virtual Cell Atlas: ARC is building the world's largest dataset of single cells to train cellular foundation models, aiming to create a "Protein Data Bank (PDB) of virtual cells" by 2030.
Whole-System Modeling: The vision extends to simulating multi-system interactions, such as the brain-gut axis, to understand complex physiological states like "runner's high" or stress-induced ulcers.
Interoception: Research is focusing on how the body signals the brain (and vice versa) to program physiology, moving beyond simple protein binding to holistic health management.
AI Doctors: Hsu predicts a future where AI integrates genotype (DNA), phenotype (biomarkers), and environmental data to provide personalized, longitudinal health recommendations rather than just reactive drug prescriptions.

Drug Development Reality: Even with perfect AI drug design, the bottleneck remains regulatory approval and clinical trials, which take years and cost hundreds of millions.
Success Rate: The pharmaceutical industry's probability of success (POS) for a drug candidate is currently around 10%; Hsu suggests AI must aim to double or triple this rate to justify its cost.
2025 Prediction: AI will enable the design of full-length antibody medicines (not just fragments) with high affinity in a "one-shot" manner.
2030 Prediction: Accurate, useful virtual cell models will mature, allowing scientists to simulate cellular behavior with high fidelity.
2050 Prediction: The development of a "scientific superintelligence" capable of fully automated, vertically integrated wet labs that can validate drug impacts in hours rather than months.
Critical Risk: Hsu warns that predictive models will fail if trained on imperfect data, noting that simulating a mouse perfectly does not guarantee accuracy for human biology due to species differences.

Closing the Feedback Loop: Current science lacks "reasoning traces" because failed experiments are often unpublished; AI can help capture this data to move science from "guess and check" to predictive modeling.
Data Utility: The most immediate enterprise value for AI in pharma currently lies in summarizing and restructuring massive regulatory documents to improve efficiency, rather than immediate molecule design.
Human Element: Success in scientific AI requires retaining human "grit" and the ability to execute projects from end-to-end, as AI cannot yet replace the final stages of project completion.