Faster Science, Better Drugs

Episode Summary

Executive Summary: Patrick Hsu and Jorge Conde argue that biology can be made dramatically faster by building virtual cell models that predict how perturbations shift cell states, similar to AlphaFold’s impact on protein structure. They link scientific slowness to incentive misalignment, fragmented disciplines, and real-world experimental bottlenecks, and say AI will matter most when it improves target selection, molecule design, and drug development outcomes.

Main Topics: ARC’s moonshot: virtual cells to accelerate science (Priority: 5/5): Patrick Hsu frames ARC’s mission as building foundation models that simulate human biology at the cellular level so experimental biology can move faster and become more scalable. Why biology is slower than AI (Priority: 5/5): The speakers attribute scientific drag to incentive structures, disciplinary fragmentation, physical distance between collaborators, and the need for real-world lab validation rather than instant GPU iteration. What a virtual cell should do (Priority: 5/5): A virtual cell is positioned as a model that predicts perturbation responses across cell states, enabling scientists to ask what changes move a cell from one state to another and to identify drug targets or combinations. From AlphaFold to a 'GPT-3 moment' in cell biology (Priority: 4/5): They compare the hoped-for inflection point in biology to AlphaFold and GPT-style capability jumps: a model that reliably predicts canonical perturbations and becomes a default tool for experimentalists. Biotech commercialization and pharma economics (Priority: 5/5): Jorge Conde explains that better science must ultimately reduce capital intensity, shorten discovery and development timelines, and increase effect sizes to improve biotech valuations and investment dynamics. Hype vs. heft in AI for biology (Priority: 4/5): The discussion distinguishes between areas with real traction, such as protein design and pathology AI, and more speculative claims like broad multimodal biological foundation models or toxicity prediction. Broader AI frontier: agents, robotics, BCI, longevity (Priority: 3/5): Conde extends his optimism beyond biotech to agents, robotics, brain-computer interfaces, and longevity, arguing that these technologies can materially improve the human experience if executed in time.

Key Arguments: Science is slow because biology requires real-world experiments, not just fast digital iteration; model development must be paired with lab-in-the-loop validation. The scientific ecosystem is hampered by incentives that favor individual papers and siloed work rather than large, cross-disciplinary, collaborative programs. Virtual cell models should focus on practical perturbation prediction: given a cell state, predict interventions that move it to another state in a useful way for wet-lab biologists. Cell biology is a sensible starting point because the cell is the fundamental unit of biological computation; once modeled well, higher levels of complexity can be built up. RNA and other scalable measurements may serve as lower-resolution mirrors of more complex layers like protein signaling, allowing useful predictions even when biology is incompletely measured. The strongest near-term AI value in biology comes from protein-related tasks, while broad multimodal biological models and toxicity-only predictors remain more speculative. Biotech improves when it can lower capital intensity, compress timelines, and increase therapeutic effect size; otherwise, even good science struggles to create strong business outcomes. The bottleneck in drug discovery is not just design but also making, testing, and clinical validation, all of which take time and are hard to fully automate. Open challenges and benchmarks can help create an AlphaFold-like public moment for virtual cells by making progress measurable and community-facing.

Data Points: Clinical trial failure rate: 90% - Used by both speakers to describe how often drugs fail in clinical trials. AlphaFold accuracy: 90%+ - Patrick cites AlphaFold as producing protein structure predictions with roughly 90% plus accuracy. Single-cell early genomics scale: 20-40 cells - Patrick notes that early single-cell sequencing papers often had only 20 or 40 cells. Perturbed single cells ARC aims to generate: 1 billion - Patrick says ARC will generate a billion perturbed single cells in the near future. GLP-1 value creation: Over $1 trillion - Patrick says the market cap added to Lilly and Novo from GLP-1 development is over a trillion dollars. Biotech market cap comparison: More than the market cap of all biotech companies combined over the last 40 years - Used to emphasize the scale of value created by GLP-1s. CASP-style competition prizes: $100,000 - ARC’s virtual cell challenge includes prizes sponsored by NVIDIA, 10x Genomics, Ultima, and others. Deep learning cycle estimate: Every 8 years - Jorge says the field seems to produce a major new deep learning shift about every eight years. Suggested model generation stage: Between GPT-1 and GPT-2 - Patrick says current biology foundation models are roughly at GPT-1 to GPT-2 capability level.

Pivotal Quotes: "I want to make science faster." — Patrick Hsu: Patrick states the core moonshot behind ARC at the start of the conversation. "Why are we so worried about modeling entire bodies over time when we can't do it for an individual cell?" — Patrick Hsu: He argues that biology should begin with the fundamental unit, the cell, before attempting whole-body simulation. "The juice needs to be worth the squeeze." — Jorge Conde: Jorge uses this phrase to explain why biotech must target high-value, high-impact diseases and outcomes.

Implications: The episode frames virtual cells as a plausible next frontier for bio-AI: if they improve target selection and perturbation prediction, they could reduce failure rates, speed discovery, and reshape pharma economics. But clinical, regulatory, and manufacturing bottlenecks will still limit how quickly AI translates into medicines.