The AI Drug Discovery Revolution
This article was created by integrating 1 related article.
Table of Contents
The Frontier of AI Drug Discovery: The Revolution Shaped by Isomorphic Labs' AlphaFold 3 and "Scientific Agents"
As artificial intelligence transforms various industries, enormous attention is being focused on its potential in healthcare — particularly drug discovery. What was once the stuff of science fiction, "AI-driven drug development," is now becoming a reality. At the forefront of this effort is Isomorphic Labs, spun out of Google DeepMind. The company pursues an ambitious vision: building a universal AI drug discovery engine that is not limited to specific diseases or targets, and has been working toward that goal since its very first day.
Last summer, the company's announcement of "AlphaFold 3" sent shockwaves through the scientific community. While the original AlphaFold revolutionized protein structure prediction, AlphaFold 3 went even further — making it possible to model with high precision not just proteins but nearly all molecular species that make up life, including DNA, RNA, and small molecules, along with their complex interactions. This achievement contributed to the Nobel Prize in Chemistry awarded to Demis Hassabis, CEO of DeepMind and founder of Isomorphic Labs.
This article, based on an interview with Max Jaderberg, Chief AI Officer (CAIO) of Isomorphic Labs, explores the company's vision for AI drug discovery, the true significance of AlphaFold 3, and the new concept of "agents for science" — delving into the full scope of their ambitious work. This is not merely a technology explainer; it is a journey into how AI is set to transform scientific research and, ultimately, human health.
From Game AI to Drug Discovery: Max Jaderberg's Path and the Potential of Reinforcement Learning The Impact of AlphaFold 3 and the Road to a "Universal Drug Discovery Engine" Data, Team, and the Future Beyond "Move 37" Conclusion
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
From Game AI to Drug Discovery: Max Jaderberg's Path and the Potential of Reinforcement Learning
Max Jaderberg's career, at the helm of AI strategy at Isomorphic Labs, has been intertwined with deep learning — particularly reinforcement learning (RL). Starting his career at DeepMind's founding as a researcher in computer vision and deep generative models, he was ultimately captivated by the world of reinforcement learning. At the time, DeepMind was a global hub for RL research, and at its center, he pursued the ultimate goal of "making AI execute any task."
What drew Jaderberg to RL was the fundamental difference in its learning paradigm. In supervised learning, which was the mainstream at the time, you give a model pairs of questions and correct answers and train it to predict the right answer. But for many real-world problems — especially exploring uncharted territory like drug discovery — the "correct answer" isn't known in advance. This is where RL's strength lies. In reinforcement learning, simply by providing feedback ("good or bad," "how good") on the actions taken by an agent (AI), the agent learns the optimal strategy through trial and error on its own. Since there's no need to teach the right answer, RL holds enormous potential for problems where humans don't yet know the solution, or for complex problems that exceed human capabilities.
Just as DeepMind early on astonished the world with its Atari game-playing, Jaderberg also used games as a testbed for AI research. He views video games as "a perfectly encapsulated, malleable world where researchers can freely manipulate, try different algorithms, and set up different scenarios." He progressed from relatively simple games like Pong and Space Invaders toward more complex games closer to real-world problems, scaling reinforcement learning algorithms along the way.
Jaderberg was particularly passionate about "zero-shot generalization" — the ability for an agent trained on a specific task to handle new, unseen tasks without additional training. Conventional RL typically trained agents from scratch for each game, but he wanted to "be able to pick up a trained agent and place it in any new task." Achieving this requires agents that generalize across the entire "task space" — they need to experience a vast variety of tasks. Rather than images or text, the training data consists of "tasks" themselves.
Attempts were made to have humans create many mini-games by hand or to procedurally generate tasks, but there were limits to the complexity achievable. That's when Jaderberg turned his attention to multiplayer games. Unlike single-player games, multiplayer games include other players. Every time the strategies and behavioral patterns of those players change, the game's character changes fundamentally, and the "tasks" the agent faces diversify infinitely. Just as chess has been played for centuries — always presenting new challenges because the opponent changes — multiplayer games can become richer and more complex task generators than anything humans could design.
Based on this insight, he developed groundbreaking AIs including "Capture the Flag" and "AlphaStar (StarCraft II)." Capture the Flag achieved human-level performance in a multiplayer first-person shooter game, dramatically advancing RL's generalization capability and multi-agent environment adaptability. StarCraft II, with its even greater complexity — enormous state spaces, imperfect information, long-term strategy — further challenged the team, and Jaderberg and colleagues rose to meet it.
The experience gained in these game AI research efforts, combined with the conviction that core concepts of deep learning (such as transformer architectures) are surprisingly transferable across different application domains, led Jaderberg to the next stage. He says, "If you point great talent, algorithms, and computing power at truly difficult problems, you can now find recipes to solve many of them." And the application domain he set his sights on was the grand frontier of drug discovery — one that could directly contribute to human health. A decade-long collaboration with Demis Hassabis, a visionary of the first order, also propelled this challenge. Hassabis's "rollout thinking" (the ability to read future developments like a chess player and make moves accordingly) and his insatiable ambition to push the boundaries of science are the driving force behind Isomorphic Labs.
The Impact of AlphaFold 3 and the Road to a "Universal Drug Discovery Engine"
The goal pursued by Isomorphic Labs extends beyond developing a single drug. The company has been pursuing an extremely ambitious vision since founding: "solving all diseases." The AI technologies they develop aim to deepen fundamental understanding of biology, and based on that knowledge, to precisely manipulate chemistry (molecular design) — dramatically enhancing the ability to modulate the biological processes that cause disease. This stands in stark contrast to traditional drug discovery approaches that specialize in specific indications or specific target proteins. What they are building is a "universal drug discovery engine" — an AI platform that can be repeatedly applied to any disease area, unconstrained by particular targets or modalities (small molecules, antibodies, etc.).
AlphaFold 3 became the landmark achievement in realizing this grand vision. While AlphaFold 2 revolutionized protein 3D structure prediction and the subsequent Multimer version enabled structural prediction of protein complexes, AlphaFold 3 went even further. It made it possible to model at atomic-level precision not just proteins but also DNA and RNA — which carry genetic information — and small molecule compounds important as pharmaceuticals, as well as how these molecular species interact with each other and form complex molecular machinery.
For example, many small molecule drugs work by binding to specific proteins in the body, modulating their function. Medicinal chemists need to design molecules that precisely bind to a target protein and produce the desired effect while minimizing side effects (such as unintended binding to other proteins). Traditionally, it has been extremely difficult to accurately determine "how a molecule binds to a protein." Experimental structure determination methods like X-ray crystallography take months to years and enormous cost, and in some cases, structure determination itself is impossible.
AlphaFold 3 dramatically alleviates this bottleneck. Chemists can now input the information about a designed molecule into a computer and visualize in 3D — within minutes to hours, with precision matching or exceeding experiment — how it interacts with a target protein or a biological molecular system including DNA/RNA. Changes to the design can be immediately simulated. This enables a dramatic shortening of the design cycle and more rational molecular design. It also opens up approaches to molecular species such as transcription factors (proteins that bind to DNA and regulate gene readout) that were previously difficult to handle as drug targets.
However, Jaderberg emphasizes that "AlphaFold 3 is only part of the story." The drug discovery process is not complete with just understanding binding structures. "About half a dozen more AlphaFold-level breakthroughs are needed," he says, to build a truly effective drug discovery engine. This means acquiring the ability to predict and understand other fundamental problems in biology and chemistry — beyond structural prediction — with experimental-level precision. For example, how strongly a molecule binds to its target (binding affinity), how binding changes protein function, how the designed molecule is absorbed, distributed, metabolized, and excreted in the body (ADME properties), and whether it shows toxicity — all of these must be considered.
Isomorphic Labs is also developing "universal" predictive models for these challenges — models that do not depend on specific targets or chemical classes, just as with AlphaFold. In their internal research projects, models that predict these critical properties with high accuracy, beyond structural prediction, are already being developed and are being utilized in actual drug discovery programs.
Even more important is the recognition that predictive models alone cannot solve the ultimate challenges of drug discovery. Jaderberg describes the size of the candidate space for druggable small molecules as "10 to the power of 60" — an astronomically large number, far exceeding the total number of atoms in the known universe. Even if there were a perfect predictive model capable of evaluating one billion (10 to the power of 9) molecules per second, only a tiny fraction of this vast chemical space could be explored.
This is where the concept of "Agents for Science" comes in — AI agents that don't just make predictions, but learn autonomously and efficiently explore the vast chemical space to discover promising molecular designs.
Core challenges in drug discovery:
Prediction: Predicting with high accuracy the structure, properties (binding affinity, ADME, toxicity, etc.), interactions with targets, and behavior in living organisms. The role filled by AlphaFold 3 and similar models.
Exploration/Search: Efficiently finding molecules with desirable properties from the vast chemical space of 10^60 candidates. Exhaustive search is impossible.
AI-based solution approaches:
- Using predictive models: Utilize high-precision predictive models as evaluation functions
- Generative models: Generate new molecular structures with potentially desirable properties
- Search algorithms/agents: Combine predictive and generative models using reinforcement learning to autonomously explore promising regions and optimize designs — similar to how AlphaGo didn't explore every possible move in Go but learned by searching for promising ones
In other words, Isomorphic Labs is simultaneously building a high-precision "map" of the biochemical world (predictive models) and training "explorers" (generative models and search agents) that use that map to efficiently traverse unknown territory and find treasure (drug candidates). Through the twin wheels of prediction and exploration, they aim to transcend the limitations of conventional drug discovery.
Data, Team, and the Future Beyond "Move 37"
Supporting Isomorphic Labs' ambitious challenge are not just cutting-edge algorithms. Data, computing power, and the team that wields them are also indispensable elements of success. Interestingly, Demis Hassabis has stated that "the biology domain is not constrained by data." This doesn't mean data is unnecessary — rather, it means there is sufficient room to make great progress using AI models with existing public data and experimentally generatable data. Jaderberg also notes that in certain modeling spaces, it is possible to make unprecedented progress using data that has been in existence for a long time.
However, this by no means implies that data collection or generation is unnecessary. On the contrary, for future breakthroughs, the generation of new data optimized for training AI models is believed to be key. "Biological data for machine learning has not yet been created in the true sense," says Jaderberg. Traditional biological experimental data was not necessarily designed to maximize AI model learning efficiency. Going forward, the use of "synthetic data" from computational methods such as physical simulations, quantum chemistry calculations, and molecular dynamics (MD) simulations, along with active learning approaches where models propose promising data that is then experimentally validated and fed back, will become important. There are also high expectations for new experimental technologies such as organoids-on-chip, which could provide high-quality in vitro data that was previously difficult to obtain and serve as alternatives or complements to animal testing. While Isomorphic Labs doesn't have its own wet lab, it actively pursues its own data generation through partnerships and external contractors.
To tackle such complex, interdisciplinary challenges, building a team that brings together diverse expertise and enables effective collaboration is essential. Since the field of AI drug discovery itself is new, people who are simultaneously "world-class experts in drug discovery" and "world-class experts in machine learning" barely exist today. The approach Isomorphic Labs has adopted is to literally "seat top experts from each domain next to each other." Rather than existing as separate teams, they interact daily, learning each other's technical vocabulary and sharing problem awareness — promoting innovation like a chemical reaction. Jaderberg says, "Many ML team members join with no prior knowledge of biology or chemistry, and that can actually become a strength." Naive questions and first-principles thinking can trigger new approaches unconstrained by existing dogma. "Being a bit naive, highly curious, and proactive" are the key qualities he believes lead to success in this field.
Alongside its internal research and development, Isomorphic Labs also places importance on returning its achievements to the scientific community. Simultaneously with the announcement of AlphaFold 3, for non-commercial academic research purposes, portions of the model's code and weights were made publicly available for free through the "AlphaFold Server." This inherits the open spirit of the AlphaFold series, intending to promote utilization in a wide range of fields beyond drug discovery — including basic biological research — and to contribute to the advancement of science as a whole.
Looking ahead, Jaderberg hints at further evolution of the AlphaFold series. The current AlphaFold 3 primarily predicts static structures, but in living organisms, molecules are constantly in motion. In the future, capturing the "dynamics (movement)" of these molecules could lead to deeper understanding and more precise design.
When will the "GPT-3 moment" in AI drug discovery arrive — the moment when what AI produces reaches or exceeds human quality? Jaderberg predicts it will manifest differently from GPT-3 in text generation — more like the "37th move" when AlphaGo defeated the world's top Go player. Move 37 was incomprehensible to the top human players at the time, appearing at first glance to be a poor move, but ultimately proved to be the decisive, creative move that secured AlphaGo's victory. Similarly, in AI drug discovery, molecular designs proposed by AI that might seem unconventional by human chemists' intuitions or existing knowledge — yet are physically and chemically sound and possess excellent properties — may increasingly appear. In fact, Jaderberg says cases are already emerging internally at Isomorphic Labs where AI-proposed designs have been shown experimentally to surpass human experts' judgments. This suggests that AI could become a "creative partner" that brings unknown scientific discoveries — going beyond being a mere tool.
It will still take time before AI-designed drug candidates actually advance to clinical trials and receive approval, but Isomorphic Labs' internal programs and partnerships are making steady progress. Jaderberg points out that as the number of AI-designed candidates increases, improving the efficiency and speed of the clinical development process itself will also become important. As AI prediction accuracy for toxicity and other properties improves, new clinical trial designs may become possible — delivering new drugs to patients more safely and rapidly in collaboration with regulatory authorities.
Ultimately, if companies like Isomorphic Labs succeed, what happens to the traditional pharmaceutical industry? Jaderberg sees convergence rather than confrontation. "In five years, conducting drug discovery without AI will be unthinkable," he says definitively. AI will become a fundamental tool in biology and chemistry — just as mathematics became the basic language of science — and the entire industry will adapt to utilize it.
Conclusion
The AI drug discovery being pursued by Isomorphic Labs aims not merely at efficiency gains or cost reduction, but at fundamentally changing the very nature of drug development. As Max Jaderberg describes, the construction of a "universal drug discovery engine" to confront all diseases — without specializing in specific conditions or targets — is a true paradigm shift. AlphaFold 3 is an important milestone in this journey, having dramatically deepened our understanding of life at the molecular level, but this is merely a prelude.
Beyond structural prediction, "half a dozen AlphaFold-level breakthroughs" in predicting multifaceted properties such as molecular function, in vivo dynamics, and toxicity — combined with "agents for science" to explore the vast chemical space of 10^60 candidates — could accelerate the discovery of drug candidates beyond anything previously imaginable. This suggests a future where AI brings discoveries that transcend human intuition and preconceptions — like AlphaGo's "Move 37."
Data strategy, interdisciplinary team building, and contributions to the scientific community are also important elements in realizing this grand vision. As AI permeates biology and chemistry as fundamental tools, the transformation of the entire pharmaceutical industry will also accelerate. Isomorphic Labs' challenge holds infinite potential for AI to contribute to human health and well-being — and its progress cannot be ignored.
Reference: https://www.youtube.com/watch?v=LrMKsBtx5Bc
TIMEWELL AI Consulting
TIMEWELL is a professional team supporting business transformation in the age of AI agents.
Services Offered
- AI Agent Implementation Support: Business automation using GPT-5.2, Claude Opus 4.5, and Gemini 3
- GEO Strategy Consulting: Content marketing strategy for the AI search era
- DX Promotion & New Business Development: Business model transformation leveraging AI
In 2026, AI is evolving from "something you use" to "something you work with." Let's think through your AI strategy together.
Related Articles
- The Reality of Working Reduced Hours After Two Maternity Leaves, and How My Perspective on Work Has Changed | TIMEWELL
- Before Taking Parental Leave (Personal Account Part 2): Three Essential Steps to Take Parental Leave Even During Busy Season
- Finding My Own Way as the Fifth Generation of a Construction Company Committed to the Field — Fujita Construction