Interview, Fireside Chat
Faster Science, Better Drugs
- Virtual cell models at ARC aim to simulate human biology to accelerate scientific progress and become the standard tool for experimentalists within a single lifetime, starting with individual cells and scaling to pairs, tissues, and intact animal environments.
- The organization plans to integrate neuroscience, immunology, machine learning, chemical biology, and genomics to increase "collision frequency," with the goal of conducting experiments at the speed of neural network forward passes and transitioning from an engineering shop to a research institute capable of inventing novel technologies.
- Model accuracy is expected to improve by layering protein information and spatial tokens on top of RNA data, eventually reaching an "AlphaFold moment" where the model predicts correct perturbations to shift cell states with 90% accuracy, while temporal dynamics are planned for future integration.
- A "GPT-3 moment" for biology is predicted to occur via a public release that could rediscover known discoveries or alter public perception, potentially leading to the generation of "a trillion binders in silico" and the development of drugs with large effect sizes for difficult diseases like obesity, neurodegeneration, and cancer.
- While physical bottlenecks in making, animal testing, and human trials are expected to remain primary constraints, AI is projected to become a native part of the drug discovery stack within a few years, compressing early discovery time more than clinical development timelines.
- The industry is forecasted to shift toward treating cancer as a chronic condition and addressing previously unaddressable genetic medicine problems, with capital intensity expected to decrease over time as technology improves and computational costs drop.
- Dario Amadei's intuition regarding multi-parallelized discovery agents is considered tangible today, with a prediction that such methods could prevent infectious diseases and double lifespans within the next decade.
- A net new deep learning architecture is expected in 2025, following an eight-year historical pattern, while computer use agents are predicted to trail coding agents by approximately one year before achieving error-free work trajectories lasting from minutes to days.
- The "Virtual Cell Challenge" is planned as an open competition where experts train models to assess capabilities over subsequent years, and new technologies targeting longevity, sleep improvement, and robotics are expected to change the world within five to eight years.
- Future progress will rely on "tech-optimist" ideas combining better target understanding, new creative medicine designs, and virtual cell integration, with investment value inflection anticipated if timelines are compressed, effect sizes increase, and capital intensity decreases.
- Investment in synthetic biology, brain-computer interfaces, and robotics (industrial and consumer) will continue to expand to improve the human experience over the coming decades, while "RL gyms" are expected to be surpassed by new learning methods citing interests in model merging and evolutionary selection.