Faster Science, Better Drugs
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
This episode dives deep into the future of science and drug discovery with Patrick Xu, co-founder of the Arc Institute, and Jorge Conde from Andreessen Horowitz. The conversation centers on Arc's ambitious moonshot to create "virtual cells" - foundation models that can simulate human biology at the cellular level. (00:00) The discussion explores why scientific progress moves so slowly, how AI could dramatically accelerate biological research, and what it would take to achieve an "AlphaFold moment" for cell biology. (00:41)
- Main Theme: Making science faster through virtual cell models and AI-powered biological simulation, addressing the complex incentive structures that slow scientific progress, and exploring how breakthrough technologies translate into real-world drugs and business outcomes.
Speakers
Patrick Xu
Patrick Xu is the co-founder of the Arc Institute and a Sixteen Z General Partner. He completed his PhD at the Broad Institute during the development of revolutionary technologies like single cell genomics, human genetics, and CRISPR gene editing. At Arc, he's leading the charge to create virtual cells and simulate human biology with foundation models, aiming to make science fundamentally faster and more efficient.
Jorge Conde
Jorge Conde is a General Partner at Andreessen Horowitz, where he leads the firm's bio practice. He focuses on early-stage investments in biotechnology and pharmaceutical companies, with particular expertise in how AI and technology can transform drug discovery and development. He brings deep industry knowledge about the bottlenecks and opportunities in bringing new medicines to market.
Key Takeaways
Science Slowness Is an Incentive Problem, Not Just a Technical One
The root cause of slow scientific progress isn't just technical limitations - it's a complex "Gordian knot" of misaligned incentives across funding, training systems, and career structures. (02:29) Traditional academic systems incentivize individual researchers to publish their own papers rather than collaborate on bigger, multidisciplinary projects. Arc Institute was designed as an organizational experiment to bring together neuroscience, immunology, machine learning, chemical biology, and genomics under one roof to increase "collision frequency" between disciplines. This approach allows teams to tackle problems that no individual research group could handle alone, demonstrating that structural innovation is as important as technical innovation.
Virtual Cells Could Be Biology's GPT Moment
The concept of "virtual cells" represents a practical approach to modeling biology that focuses on perturbation prediction rather than complete biological simulation. (10:51) Just as AlphaFold solved protein folding with 90% accuracy without simulating all the underlying biophysics, virtual cell models aim to predict how cells move between different states when perturbed. The goal is to create a "universal cell space" where researchers can predict what perturbations are needed to move cells from diseased states to healthy ones. This could revolutionize drug discovery by enabling researchers to computationally design combination therapies and identify novel drug targets before expensive lab experiments.
RNA Data Serves as a Scalable Mirror for Complex Biology
While we can't measure everything happening in a cell, RNA expression data acts as a "lower resolution mirror" that reflects protein-level and metabolic activity. (09:07) Although RNA doesn't perfectly represent protein function, at massive data scales, transcriptional states begin to echo what's happening at the protein level. This scaling approach allows researchers to bet on technologies that are "scale ready" today while gradually layering on additional data types like spatial and temporal information. The key insight is that you don't need perfect biological measurements - you need scalable ones that capture enough signal to train useful models.
The Industry Needs Higher Effect Sizes, Not Just Better Technology
The biotechnology industry's fundamental challenge isn't just scientific - it's economic. (26:57) With 90% of drugs failing in clinical trials and massive capital intensity, the industry struggles with poor risk-adjusted returns. The success of GLP-1 drugs like Ozempic, which added over a trillion dollars in market cap to Lilly and Novo Nordisk, demonstrates the value of tackling large patient populations rather than just well-validated but small markets. The industry needs to reduce capital intensity, compress development timelines, and increase effect sizes by going after bigger, more impactful diseases rather than playing it safe with small patient populations.
AI Progress in Biology Lags Because We Don't Speak Biology Natively
AI has progressed faster in language and image generation because humans can naturally evaluate the outputs - we already know how to speak and see. (05:52) In contrast, "we don't speak the language of biology" or at best speak it "with an incredibly thick accent." This creates a fundamental bottleneck where researchers must constantly run lab-in-the-loop experiments to validate AI predictions, slowing the iteration cycle. The challenge isn't just technical - it's interpretive. Building useful AI for biology requires developing new ways to interpret "weird fuzzy outputs" from biological models and bridging the gap between computational predictions and experimental validation.
Statistics & Facts
- Patrick Xu mentioned that early 2010s single cell sequencing papers would analyze 20-40 cells, while Arc Institute plans to generate a billion perturbed single cells in the near future - demonstrating exponential scaling in biological data generation. (25:23)
- Jorge Conde noted that 90% of drugs fail in clinical trials, highlighting the massive failure rate that drives the industry's capital intensity and risk management challenges. (23:59)
- The combined market cap added to Eli Lilly and Novo Nordisk from GLP-1 development exceeds a trillion dollars, which is more than the total market cap of all biotech companies started in the last 40 years combined. (30:43)
