This episode reveals how foundation models are being built to simulate human biology, tackling the fundamental bottlenecks that could unlock a new era of AI-driven drug discovery and change the economics of the entire pharmaceutical industry.
The Mission to Accelerate Science
- Patrick, a researcher and investor, opens by stating his direct goal: to make science faster. He argues that unlike pure AI research, which iterates at the speed of GPUs, biological science is constrained by the physical world—growing cells, tissues, and animals takes time. The core mission at the Arc Institute is to build "virtual cells" using foundation models to simulate human biology, creating a default tool for experimentalists to accelerate and parallelize their work.
- Actionable Insight: The bottleneck in bio-AI is not just computation but the physical "lab-in-the-loop" validation process. Investors should look for companies that are not just building models but also integrating or innovating on high-throughput, automated lab infrastructure to close this feedback loop faster.
Why Science is Slow: A Gordian Knot of Incentives
- The conversation identifies the slow pace of scientific progress as a multifactorial problem rooted in misaligned incentives. Patrick points out that the current academic system doesn't adequately reward collaboration. Researchers are driven to publish their own papers and make individual discoveries, discouraging the large-scale, multidisciplinary efforts required to solve complex problems.
- The Arc Institute's Hypothesis: Arc was designed as an "organizational experiment" to solve this. By bringing experts in neuroscience, immunology, machine learning, and genomics under one physical roof, the goal is to "increase the collision frequency" between disciplines and foster work on large, flagship projects that no single group could tackle alone.
The Unique Challenge of Modeling Biology
- Patrick offers a candid take on why AI has progressed faster in language and images than in biology: "To be honest, it's a lot easier." He explains that humans have an intuitive, native ability to evaluate the output of a language or image model. In contrast, we don't "speak the language of biology" natively.
- The Language Barrier: When training a DNA foundation model, researchers lack an innate sense of whether the output is correct. This necessitates a slow, expensive "lab-in-the-loop" process where model predictions must be validated with physical experiments.
- Incomplete Data: The models are being built even though we cannot yet measure all components of a cell. The strategy is to bet on data that can be scaled today, like transcriptional data (RNA), using it as a lower-resolution "mirror" for more complex protein-level activity. As measurement technologies improve, more data layers (spatial, temporal) can be added.
Deconstructing the "Virtual Cell" Ambition
- The ultimate goal is to create an "AlphaFold moment" for cell biology. AlphaFold is a landmark AI model from DeepMind that accurately predicts a protein's 3D structure from its amino acid sequence. The equivalent for a virtual cell would be a model that can reliably predict how a cell's state will change in response to specific interventions.
- Perturbation Prediction: At Arc, this is operationalized as "perturbation prediction." The model would learn a universal map of cell states (e.g., healthy, inflamed, stressed) and then predict the precise combination of interventions (the "perturbations") needed to move a cell from a diseased state to a healthy one.
- A Co-Pilot for Biologists: The model is not intended to be a purely theoretical tool. It is designed to be a practical co-pilot for a wet lab biologist, suggesting concrete experiments to run, thereby creating a tight feedback loop between in silico prediction and experimental validation.
The Path to a "GPT-3 Moment" in Biology
- Patrick frames the progress of virtual cell models in terms of GPT generations, estimating the field is currently between "GPT-1 and GPT-2." The "GPT-3 moment"—a public breakthrough that changes perceptions of what's possible—will likely involve the model recapitulating famous, Nobel Prize-winning biological discoveries from scratch.
- Key Benchmark: A major test would be if the model could independently predict that the four "Yamanaka factors" can reprogram a skin cell (fibroblast) back into a stem cell-like state, a discovery that won the Nobel Prize.
- Cory's Perspective: Cory, an investor, emphasizes that the industry's biggest bottleneck remains clinical trials. While AI can compress discovery time, it has not yet significantly accelerated the process of proving a drug is safe and effective in humans.
Navigating Industry Bottlenecks and Business Models
- The discussion shifts to the economic realities of the biotech and pharma industries. Patrick notes that 90% of drugs fail in clinical trials, either because the biological target was wrong or the drug molecule itself was ineffective. This high failure rate creates a risk-averse culture.
- Cory's Analysis on Industry Health: Cory outlines three key factors needed to fix the industry's economics:
- 1. Reduce Capital Intensity: Higher success rates from better models should lower the overall capital needed.
- 2. Compress Timelines: AI is helping in early discovery, but the major bottleneck of clinical development remains.
- 3. Increase Effect Size: Better drugs targeting the right mechanisms should produce clearer, more dramatic results, making trials shorter and more decisive.
The Transformative Impact of GLP-1s and Future Ambition
- The success of GLP-1 drugs (like Ozempic) is highlighted as a pivotal moment. Patrick observes that the market cap added to companies like Eli Lilly and Novo Nordisk from these drugs "is more than the market cap of all biotech companies combined over the last 40 years."
- Strategic Implication: This massive value creation from targeting a large patient population (obesity and diabetes) has culturally increased the ambition of the entire industry. It serves as a powerful counter-narrative to the previous focus on rare diseases with small patient populations, which was seen as a way to manage risk.
Future Bottlenecks: Beyond Design to Manufacturing and Testing
- Even if AI models can perfectly design trillions of drug candidates in silico, Patrick argues that the real-world bottlenecks of manufacturing and testing will remain. He draws a compelling contrast between the US and China.
- > "China is an engineering state... Whereas, I think from, you know, the first 13 American presidents, 10 of them practice law."
- The Regulatory Hurdle: This cultural difference is reflected in the FDA's regulatory regime, which remains a major bottleneck. Solving the computational design problem does not solve the physical world challenges of producing molecules and running human trials.
AI in Drug Discovery: Separating Hype from Heft
- When asked to separate hype from reality in AI for drug discovery, Patrick provides a clear framework:
- Hype: Toxicity prediction models. The idea that an AI can reliably predict if a novel molecule will be toxic is currently overhyped.
- Heft (Real Impact): Anything related to proteins. This includes predicting protein binding, designing new proteins, and using AI to automate the work of pathologists and radiologists.
- Hope: Multimodal biological models that integrate different layers of data (genomic, spatial, etc.) to create a more holistic simulation.
The Broader AI Landscape: Agents, Architectures, and Untapped Research
- Looking beyond biology, Patrick discusses the broader AI investment landscape. He is focused on areas that "improve the human experience," including synthetic biology, brain-computer interfaces (BCIs), and robotics.
- The Next Architectural Shift: He notes that the industry is still heavily reliant on the Transformer architecture from 2017 and is "really overdue for some net new architecture." He believes many powerful ideas from academic papers published between 2009-2015, which were computationally infeasible then, can now be scaled with modern compute, creating opportunities for new foundation model labs.
- Actionable Insight for Researchers: There is immense value in revisiting and scaling older, under-cited machine learning research papers. As the cost of compute falls, ideas that failed at a small scale may now exhibit powerful scaling laws, unlocking new capabilities beyond the current Transformer-centric paradigm.
Conclusion
- This discussion highlights that while AI is poised to revolutionize biological discovery, the primary bottlenecks are no longer just computational but physical and regulatory. For investors and researchers, the key is to identify ventures that holistically address the entire cycle—from model development to automated lab validation and navigating clinical pathways.
