a16z Podcast – “Faster Science, Better Drugs”
September 15, 2025
Host: Andreessen Horowitz | Guests: Patrick Hsu (Co-founder, ARC Institute), Jorge Conde (a16z General Partner)

Episode Overview

This episode dives deep into the current frontiers and bottlenecks in biological science and drug discovery, focusing on how advanced AI and interdisciplinary collaboration could fundamentally accelerate progress. The conversation explores the moonshot goal of simulating “virtual cells,” the translation of breakthroughs into real medical and business results, and the incentive structures slowing science. The guests also reflect on what a transformative “AlphaFold moment” would look like for cell biology, and what’s truly required to compress the painfully slow timelines of drug development.

Key Discussion Points & Insights

1. The Vision: Making Science Faster with Virtual Cells

• Patrick Hsu’s Moonshot: Build “virtual cells” at ARC Institute using foundation models to simulate human biology. Goal: Make these models useful tools for experimentalists, not just for ML benchmarks.

  “Our moonshot is really to make virtual cells at ARC and simulate human biology with foundation models.” (A, 00:00)

• The challenge: Current modeling can’t reliably capture even a single cell, so modeling entire bodies is still a distant goal.
• Importance: Compressing experimental timescales—doing in silico experiments at the “speed of a forward pass” in a model—could revolutionize drug discovery.

2. Why Is Science So Slow?

• Incentive Knots and Multidisciplinarity:
  • Science is slowed by complex incentive structures: publication pressure, career competition, and separation between basic and commercial science.
  • Universities, despite the multidisciplinary premise, are physically and bureaucratically siloed. ARC Institute is in part an experiment in high “collision frequency” — integrating neuroscience, immunology, machine learning, chemical biology, and genomics in one location.

    “It’s this weird Gordian knot that ultimately comes down to incentives.” (A, 02:30)

• Technology vs. Biology Complexity:
  • AI progresses faster in domains (images, language) we “natively” understand and can easily interpret, unlike biology.

    “Technology is easier than biology. Natural language and video modeling is easier than modeling biology.” (A, 05:49)

3. Virtual Cells: Concept & Path to AlphaFold Moment

• Definition & Ambition:
  • A “virtual cell” model would predict the outcomes of perturbations—such as drugs or genetic edits—in different cell types and states.
  • Analogous to AlphaFold for proteins: Move from theory to a practically useful tool for biologists.

  “We kind of want to get to that point with virtual cells as well... to do perturbation prediction.” (A, 11:16)

• Stages of Complexity:
  • From modeling individual cells → pairs → tissues → entire bodies, each stage layering more complexity.
  • Focus on perturbation prediction as a practical core for wet-lab researchers—a model that can suggest what 12 experiments a biologist should run, not just theoretical outputs.
• Data Limitations and Representation:
  • Current models rely on what’s scalable now: transcriptomics (RNA), with slower progress in proteomics and spatial data.

  “You have to bet on what you can scale today.” (A, 09:51)

  • RNA is a “lower-resolution mirror” for protein-level biology, but with enough data, useful predictions can emerge.

4. From AI Breakthroughs to New Drugs: Bottlenecks

• Why hasn't an alpha-level ML tool translated into blockbuster drug discovery yet?
  • AI gains in simulation don’t yet transfer cleanly through “making” and “testing” bottlenecks—real-world experiments and clinical trials remain slow, expensive, and highly regulated.

    “There’s designing, there’s the making, there’s the testing, there’s the approvals... testing takes real hours, days, months, years.” (A, 41:04)

• Drug Development Challenges:
  • 90% failure rate in clinical trials—much due to targeting the wrong mechanism or poor molecular design.
  • Even with transformative AI models, physical and regulatory bottlenecks (e.g., clinical testing timelines) won’t disappear soon.

    “That bottleneck should exist. I’m not suggesting we’ve got to remove it. But are there ways to reduce the cost and time associated with getting through the bottleneck of human clinical trials?” (B, 27:06)

5. Bringing AI to Biology—Where’s the Hype, the Hope, and the Heft?

• Hype:
  • Toxicity prediction models—claims often outpace reality.
  • Broad, “multimodal” biological models are still immature.
• Heft:
  • Protein structure design (thanks to AlphaFold).
  • Increasingly, protein design and AI for pathology (e.g., aiding radiologists).
• Hope:
  • “Virtual cell” and perturbation prediction promise real industrial utility and could enable a new era of AI-native pharma.

  “We believe in virtual cells not just because we think it will be a fountain of fundamental mechanistic insights... it could be industrially really useful.” (A, 23:32)

6. Business & Industry Impact: Capital, Risk, and Value Creation

• Industry Trends:
  • For most of biotech history, value creation has clustered around safe, low-risk, small populations (“low-hanging fruit”).
  • Successful “big effect size” drugs (e.g., GLP-1s) now inspire bolder ambitions, proving the societal and economic impact of cracking complex, high-prevalence diseases.

  “The market cap added to Lilly and Novo... more than the market cap of all biotech companies combined over the last 40 years.” (A, 31:01)

• Capital intensity—both in model training and clinical testing—remains a central challenge for startups and investors.

  “If the capital intensity goes down and the value creation goes up, it becomes easier to invest in these companies in the early days because you get rewarded for coming in early.” (B, 29:03)

7. What Would a “GPT-3 Moment” Look Like for Biology?

• Public model release that “shocks the public” with new capabilities, inspires talent influx, and demonstrates “textbook” biological predictions (i.e., rediscovering Nobel-level discoveries like the Yamanaka factors for cell reprogramming).

  “Could it predict that the four Yamanaka factors would reprogram the fibroblast into a stem like state... something that won the Nobel Prize in 2009?” (A, 17:33)

• The ultimate vision: an AI pharma company powered not just by buzzwords, but by step-change improvements in target ID and drug design—reducing both time and risk.

8. The Limits of Simulation vs. Understanding

• Simulation may predict outcomes without explaining the underlying mechanisms (mirroring debates in weather modeling and AlphaFold).

  “If you want to predict the weather, right, you can build AI models that will predict whether or not it will rain next Tuesday... Similarly with a virtual cell model, it may not tell me literally why...” (A, 45:00)

9. What's Next for ARC and the Field?

• ARC’s Virtual Cell Challenge: An open global competition to catalyze progress on virtual cell models, akin to how AlphaFold built on the CASP protein folding competition. Prize pool sponsored by industry leaders.

  “We created our own virtual cell challenge... an open competition that anyone can enter where you can train perturbation prediction models and we can openly and transparently assess these model capabilities.” (A, 54:34)

Notable Quotes & Memorable Moments (With Timestamps)

• On the vision of virtual cells:

  “If we can figure out how to model the fundamental unit of biology, the cell, then from that we should be able to build.” (B, 00:14)

• On complexity and incentives:

  “Science is slow…it's this weird Gordian knot that comes down to incentives.” (A, 02:30)

• On technology vs biology:

  “Technology is easier than biology…we don’t speak the language of biology. At very best, with an incredibly thick accent.” (A, 05:49–06:18)

• On AlphaFold and the leap to virtual cells:

  “Anytime you want to, you know, work with a protein, if you don’t have an experimentally solved structure, you’re just going to fold it with this algorithm. We kind of want to get to that point with virtual cells as well.” (A, 11:16)

• On scaling data and layers of understanding:

  “RNA representation is a mirror…it might be a lower resolution mirror for what’s happening at the protein layer, but eventually…at some sort of mirror echo.” (A, 09:31)

• On bottlenecks in drug development:

  “That bottleneck should exist. I’m not suggesting we’ve got to remove it. But are there ways to reduce the cost and time associated with getting through…human clinical trials?” (B, 27:06)

• On AI for drug design’s future ubiquity:

  “In just a few years, [AI for drugs] will just be a native part of the stack. Just like we use the Internet and phones…we’re going to have AI in all parts of the stack.” (A, 40:50)

• On hope for the field:

  “If we could do that, we could make a new AI, like vertically integrated AI enabled pharma company.” (A, 14:38)

• On making virtual cells a reality:

  “I just, I want this thing to exist in the world…I'd just be happy that someone does it.” (A, 55:12)

Timestamps for Key Segments

• 00:00–03:35 — The moonshot: Virtual cells & faster science
• 03:35–05:20 — Organizational and incentive barriers to science
• 05:31–07:00 — Why tech moves faster than biology
• 10:47–14:43 — What is a virtual cell; ambition and bottlenecks explained
• 15:09–19:43 — How far from an “AlphaFold moment” for cells?
• 22:18–23:32 — Why start with cells, not bodies?
• 26:11–30:31 — Investment and business implications; capital intensity
• 31:00–34:08 — Lessons from GLP-1s, industry ambition
• 36:11–38:38 — The hard, slow loop from in silico to real, tested drugs
• 39:26–42:26 — Hype vs. hope vs. heft in current AI/biotech
• 46:30–49:50 — Where are AI investments moving? Opportunities and overhyped areas
• 54:34–55:33 — Announcing ARC’s Virtual Cell Challenge

Conclusion & Opportunities

This episode offers an unvarnished look at both the promise and practical hurdles of integrating AI into the biology and drug discovery pipeline. While massive modeling advances—like AlphaFold for proteins—signal what’s possible, the stubborn pace of progress in clinical development means the next big breakthrough will require both technical and organizational innovation. ARC’s Virtual Cell Challenge and similar efforts aim to catalyze this transition from dream to reality—and listeners are invited to join the mission.

To learn more or enter the Virtual Cell Challenge, visit: virtualcellchallenge.org