Faster Science, Better Drugs
a16z PodcastFull Title
Faster Science, Better Drugs
Summary
The episode explores how to accelerate scientific discovery, particularly in drug development, by leveraging foundation models and creating "virtual cells."
It discusses the challenges of traditional scientific research, the potential of AI in biology, and the future of biotech and pharma business models.
Key Points
- The ARC Institute aims to accelerate science by creating virtual cells and simulating human biology with foundation models, tackling the fundamental unit of life.
- Traditional science is slow due to a "Gordian knot" of incentives, training systems, and the increasing multidisciplinary nature of research, making it hard for individual groups to cover multiple domains.
- ARC's organizational experiment gathers diverse fields like neuroscience, immunology, machine learning, and genomics under one roof to increase interdisciplinary "collision frequency."
- AI has progressed faster in language and image generation than biology because understanding and evaluating outputs in biology is more complex and less intuitive than with human language or visuals.
- Creating accurate biological simulators like virtual cells is challenging because we don't fully understand all cellular components and their functions, requiring a "lab in the loop" for validation.
- The "AlphaFold moment" for cell biology would mean having a default, accurate tool for cell biology predictions, analogous to AlphaFold's impact on protein structure prediction.
- Virtual cell models are being developed to predict perturbation responses, aiming to guide experimentalists in labs by suggesting interventions to shift cells between different states.
- The ambition is to create a vertically integrated, AI-enabled pharma company that can discover new drug targets and design effective drug compositions.
- The biotech and pharma industries face challenges with high capital intensity and long timelines, with a significant percentage of drugs failing in clinical trials.
- Accelerating drug discovery requires improving success rates, compressing timelines, and increasing effect sizes, which technological advancements like AI can facilitate.
- The development of GLP-1 drugs has demonstrated the immense value that can be created by addressing large patient populations and significant societal problems.
- While AI can help design molecules and predict binding, the physical process of making and testing drugs through preclinical and clinical trials remains a significant bottleneck.
- Future breakthroughs in synthetic biology, brain-computer interfaces, and robotics are expected to fundamentally improve the human experience.
- The current AI architecture is based on 2017 research, and there's anticipation for new architectures and the scaling of previously overlooked academic ideas with decreasing compute costs.
- Agents that perform "real work" by replacing human productivity, such as in coding, legal, BPO, and medicine, are seen as having significant heft, despite current errors.
Conclusion
Accelerating scientific discovery, particularly in drug development, is achievable through interdisciplinary collaboration, advanced AI models, and innovative organizational structures like the ARC Institute.
The complexity of biological systems means that AI's impact will be gradual, requiring continuous integration into every stage of the drug discovery and development pipeline.
Future advancements in synthetic biology, BCIs, and robotics, supported by effective AI and investment, hold the potential to significantly improve human life and address major societal challenges.
Discussion Topics
- How can interdisciplinary collaboration be effectively fostered to break down existing incentive structures that slow down scientific progress?
- What ethical considerations and validation methods are crucial as AI models become more integral to drug discovery and biological research?
- Beyond drug development, what are the most promising applications of advanced AI and computational modeling in improving human health and longevity?
Key Terms
- Foundation Models
- Large AI models trained on vast amounts of data that can be adapted for various downstream tasks, such as in biology.
- Virtual Cells
- Computational models designed to simulate the behavior and function of individual cells, aiming to predict biological responses.
- Perturbation Prediction
- Using models to predict how a biological system (like a cell) will respond to specific interventions or changes.
- AlphaFold
- An AI system developed by DeepMind that predicts protein structures from their amino acid sequences with high accuracy.
- Multidisciplinary
- Involving or representing several academic disciplines or professional specializations.
- Gordian Knot
- A complex problem that is difficult to solve.
- Collision Frequency
- In a scientific context, this refers to the rate at which different ideas or researchers from various disciplines interact and combine.
- Polypharmacology
- The phenomenon where a single drug molecule interacts with multiple biological targets.
- Digital Twins
- Virtual replicas of physical objects, processes, or systems used for simulation, monitoring, and optimization.
- Ground Truth
- The actual, real-world facts or data against which a model's predictions are compared and validated.
- GBT-1/2/3/4/5
- Refers to successive generations of Generative Pre-trained Transformer models, indicating increasing capabilities.
- RNA
- Ribonucleic acid, a molecule involved in protein synthesis and other cellular functions.
- Protein Folding
- The physical process by which a protein chain acquires its native three-dimensional structure.
- GLP-1
- Glucagon-like peptide-1, a hormone that stimulates insulin secretion and has applications in treating diabetes and obesity.
- BCI
- Brain-Computer Interface, a technology that allows direct communication pathways between the brain and an external device.
Timeline
The ARC Institute's moonshot is to make science faster by creating virtual cells and simulating human biology with foundation models.
Science is slow due to a complex interplay of incentives, training systems, and the increasing multidisciplinary nature of research.
ARC was built as an organizational experiment to foster interdisciplinary collaboration by bringing diverse research fields together.
AI has progressed faster in image and language generation than biology due to the inherent complexity of interpreting biological data.
Iteration cycles in biology are slowed by the need for experimental validation to test model predictions.
The virtual cell concept aims to be a practical co-pilot for biologists, guiding experimental decisions rather than just theoretical exploration.
AlphaFold is a prime example of ML's success in biology, solving the protein folding problem and serving as a benchmark for what a "virtual cell moment" could look like.
ARC operationalizes virtual cell models through perturbation prediction, aiming to map and manipulate cell states computationally.
The goal is to move beyond accidental polypharmacology to purposeful manipulation of cell states to treat complex diseases.
The aim is to create a vertically integrated, AI-enabled pharma company through fundamental research capability breakthroughs.
The virtual cell "AlphaFold moment" would involve models accurately predicting perturbations that shift cells between states 90% of the time.
Current DNA foundation models are like "blurry pictures of life," with ongoing research to improve their accuracy and capability.
A key validation for virtual cell models would be rediscovering known biological discoveries, such as the Yamanaka factors for stem cell reprogramming.
Textbooks are compressed representations of knowledge, and future research may uncover exceptions and new insights beyond current understanding.
The term "virtual cells" is used to describe the goal of modeling biology, even if more abstract concepts like "digital twins" are more readily accepted.
The complexity of modeling entire bodies stems from the difficulty in fully understanding and modeling individual cells.
Biotech and pharma growth relies on innovations translating into effective business models, with AI needing to compete for R&D budgets, not just SaaS.
The industry's bottlenecks are the necessary stages of drug development, particularly human clinical trials, which are hard to significantly speed up.
Improving the industry requires reducing capital intensity, compressing timelines, and increasing the effect size of drugs.
The massive market cap increase for GLP-1s highlights the value of addressing large patient populations and complex societal problems.
Future breakthroughs will likely involve a combination of better understanding what to target, designing medicines with new modalities, and using virtual cell models.
The physical processes of making and testing drugs remain a significant bottleneck, even with AI-driven design.
AlphaFold solved protein folding, but drug discovery involves more complex biological systems and requires multi-stage testing.
AI hype exists in toxicity prediction and multimodal biological models, while there's heft in protein-related AI and pathology predictions.
AI has not yet turned out drugs because drug development is a long, multi-faceted process, but AI is becoming a native part of the entire stack.
Dario Amadei's essay predicted that independent scientific discoveries could be multi-parallelized with sufficiently predictive models.
It's almost certain that current biological data is incomplete, with AI models needing to infer missing information.
There's a distinction between mechanistic models and meteorological simulation-type models; virtual cells aim for the latter initially.
Jorge Conde's investment focus is on improving the human experience through synthetic biology, brain-computer interfaces, and robotics.
Successful technological development requires a combination of technical innovation, product intuition, and business acumen, funded at the right time.
Agents that perform "real work" and impact the services economy are seen as having significant heft, with hype surrounding model capabilities.
New AI architectures are expected around 2025, and scaling academic ideas with decreasing compute costs will create new opportunities.
The ARC Institute is hosting a Virtual Cell Challenge to drive innovation and transparently assess model capabilities.
Episode Details
- Podcast
- a16z Podcast
- Episode
- Faster Science, Better Drugs
- Official Link
- https://a16z.com/podcasts/a16z-podcast/
- Published
- September 15, 2025