"I want to make science faster," declared Patrick Hsu, co-founder of Arc Institute, articulating a sentiment that resonates deeply within the innovation ecosystem. His audacious moonshot, unveiled during a compelling a16z podcast interview with Erik Torenberg and Jorge Conde, centers on the development of "virtual cells" to "simulate human biology with foundation models." This ambitious endeavor seeks to fundamentally transform scientific discovery, particularly in drug development, by addressing the systemic inefficiencies that have long plagued biological research.
Hsu, joined by a16z General Partner Jorge Conde, delved into the multifaceted reasons behind the sluggish pace of scientific progress. Conde provocatively questioned, "Whose fault is that?" Hsu's response highlighted a "weird Gordian knot" of factors, notably misaligned "incentives" within academic and industrial structures, and a fragmented "training system." Traditional research models, often siloed, struggle with the increasing interdisciplinarity required for breakthroughs. It's difficult for any single research group or company to excel across five distinct domains—neuroscience, immunology, machine learning, chemical biology, and genomics—simultaneously. Arc Institute, Hsu explained, is designed as an "organizational experiment" to bring these diverse disciplines under one physical roof, fostering "collision frequency" and enabling work on "bigger flagship projects" that transcend individual capabilities.
The conversation naturally pivoted to the role of artificial intelligence, questioning why AI has demonstrated such accelerated progress in fields like natural language processing and image generation compared to biology. Hsu offered a sharp insight: "Natural language and video modeling is easier than modeling biology." Humans possess an inherent, intuitive understanding of language and visual data, allowing for native evaluation of AI outputs. In biology, however, the "language of biology" is not natively spoken by humans. We don't inherently understand DNA or protein interactions. This necessitates a "lab in the loop" approach, where AI's "weird fuzzy outputs" must be validated through real-world experiments, significantly slowing the iteration cycle.
The vision for virtual cells aims to overcome this fundamental hurdle. The goal is to achieve an "AlphaFold moment" for cell biology, mirroring the groundbreaking success in protein folding prediction. AlphaFold, while not perfect, provides protein structures with over 90% accuracy, dramatically accelerating research by giving scientists a highly probable end-state. Similarly, virtual cells would enable "perturbation prediction," allowing researchers to computationally model how cells respond to various interventions. This capability is critical for drug discovery, moving from an accidental "this thing happens to have a whole bunch of different targets" approach to a purposeful, combinatorial manipulation of cell states.
Hsu outlined a tiered approach, distinguishing between invention, engineering, and scaling. While some biotechnological aspects are "scale-ready," many require novel invention. Arc is investing in this invention, creating "blurry pictures of life" from genomic data, which over time, through successive generations of models and data, will become clearer, more accurate representations. The ability to perform "in silico target ID" – identifying new drug targets and the precise drug compositions needed to achieve desired cellular changes – represents a paradigm shift. This would empower a new generation of "vertically integrated AI-enabled pharma companies," vastly improving the industry's efficiency.
The current biotech and pharma landscape is characterized by significant capital intensity and high failure rates in clinical trials. Over 90% of drugs fail in clinical trials, often because "we're going after the wrong thing" or "we made the wrong thing to go after the right thing," as Conde aptly summarized. The emergence of highly successful drugs like GLP-1s, which have added over a trillion dollars in market capitalization to companies like Eli Lilly and Novo Nordisk, demonstrates the immense value created when scientific breakthroughs address widespread societal health problems. This success is not merely a financial triumph but a testament to the power of targeted, effective intervention.
The industry desperately needs to increase its "hit rate" and "effect size" for new drugs. This involves not only improving early discovery but also streamlining the arduous and costly clinical trial process. While regulatory bottlenecks are a necessary safeguard, innovations in trial design and patient enrollment could compress timelines. Hsu emphasized that the true breakthrough will come when AI models can reliably predict perturbations, suggesting specific interventions that would shift a cell from a diseased state to a healthy one. This would not just accelerate the process but fundamentally de-risk it, making the "journey worth it" for every drug developed.
