The Frontier of AI Drug Discovery: Isomorphic Labs, AlphaFold 3, and the Science Agent Revolution

From Hamamoto at TIMEWELL

This is Hamamoto from TIMEWELL Corporation.

AI Drug Discovery Is No Longer Science Fiction

As AI reshapes industries across the board, healthcare — and drug discovery in particular — has attracted extraordinary attention. What once seemed like science fiction is becoming operational reality. At the leading edge of this transformation is Isomorphic Labs, a company spun out of Google DeepMind, pursuing one of the most ambitious goals in modern science: building a universal AI engine capable of developing treatments for all diseases.

Last year's release of AlphaFold 3 sent a wave through the scientific community. Where AlphaFold 2 had revolutionized protein structure prediction, AlphaFold 3 extended that capability to model interactions between virtually all major biological molecules — proteins, DNA, RNA, and small-molecule compounds — at atomic-level precision. This achievement contributed to DeepMind CEO and Isomorphic Labs founder Demis Hassabis receiving the Nobel Prize in Chemistry.

This article, drawing on an interview with Isomorphic Labs Chief AI Officer Max Jaderberg, examines the company's vision for AI-driven drug discovery, the true significance of AlphaFold 3, and what "Agents for Science" means in practice.

• From Game AI to Drug Discovery: Max Jaderberg's Path and the Promise of Reinforcement Learning
• AlphaFold 3 and the Road to a Universal Drug Discovery Engine
• Data, Teams, and the "Move 37" Moment That Lies Ahead
• Summary

From Game AI to Drug Discovery

Max Jaderberg's career has been defined by deep learning — and specifically by reinforcement learning (RL). As an early DeepMind team member, he progressed from computer vision and generative model research toward RL, which became his primary domain. DeepMind at that time was a world center of RL research, and at its core Jaderberg pursued the ultimate goal: enabling AI to perform any task.

His attraction to RL was fundamental. Supervised learning works by providing the model with question-answer pairs and training it to predict correct outputs. But in many real-world problems — especially in drug discovery — there is no known correct answer. RL resolves this: an agent takes actions and receives feedback indicating how good or bad the outcome was, allowing it to discover optimal strategies through trial and error without being told the solution. This makes RL particularly suited to problems where humans have not yet found answers, and where the solution space exceeds human cognitive reach.

Just as DeepMind famously used Atari games to demonstrate early RL capabilities, Jaderberg used games as test environments — "perfectly encapsulated, malleable worlds that researchers can manipulate freely." He progressed from simpler games toward richer environments that more closely resembled real-world complexity.

His particular focus was "zero-shot generalization" — the ability of an agent trained on one set of tasks to handle entirely novel tasks without additional training. Conventional RL trained separate agents for each game; Jaderberg wanted a single agent that could generalize across the full task space. Achieving this required massive diversity of training tasks — not images or text, but tasks themselves as the training data.

Human-generated mini-games and procedural task generation hit complexity ceilings. The breakthrough insight: multiplayer games. Unlike single-player environments, multiplayer games have other agents whose changing strategies fundamentally alter the nature of the game — generating effectively infinite task variation. Just as chess has been played for centuries with every new opponent creating a new challenge, multiplayer games are far richer task generators than any human-designed curriculum.

This led to landmark AI systems: "Capture the Flag" (multiplayer FPS at human-level performance) and "AlphaStar" (StarCraft II). StarCraft II presented vastly expanded state spaces, imperfect information, and long-horizon strategic planning — and the team solved it.

The lesson Jaderberg drew: the core architectural concepts of deep learning (transformers, for instance) transfer with surprising effectiveness across application domains. With great people, strong algorithms, and sufficient compute, many previously intractable problems now have a recipe for solution. Drug discovery — a domain with direct impact on human health — became the natural next frontier.

AlphaFold 3 and the Universal Drug Discovery Engine

Isomorphic Labs' goal is not to develop any single drug. The ambition, from day one, has been to solve all diseases — by building a general-purpose AI engine that deepens biological understanding and uses that understanding to precisely design molecular interventions.

AlphaFold 3 is the most prominent milestone on this path. AlphaFold 2 had transformed protein structure prediction; AlphaFold 3 went further, modeling how proteins interact with DNA, RNA, and small-molecule drugs across complex biological systems — at atomic-level precision, in minutes rather than months.

The practical significance for drug discovery is substantial. Most small-molecule drugs work by binding to specific proteins and modulating their function. Designing effective drugs requires understanding precisely how a candidate molecule binds to its target — including whether it might unintentionally bind to other proteins and cause side effects. Traditionally, determining binding structures required X-ray crystallography or similar techniques, often taking months or years and sometimes failing entirely.

AlphaFold 3 removes this bottleneck. A chemist can now input a molecule's structure and see — in hours rather than months — how it interacts with the target protein and the surrounding molecular system. Design changes can be simulated immediately. Molecular classes previously considered "undruggable," such as transcription factors that regulate gene expression by binding to DNA, become more tractable.

Yet Jaderberg is careful to note that "AlphaFold 3 is only one part of the story." Drug discovery requires far more than binding structure prediction. He estimates that "another half-dozen AlphaFold-level breakthroughs" remain necessary — across binding affinity, protein functional change upon binding, ADME properties (absorption, distribution, metabolism, excretion), toxicity, and more. Isomorphic Labs is developing generalized predictive models for all of these properties, and several are already in active use within the company's drug programs.

The deeper challenge: even perfect predictive models cannot solve drug discovery alone. The chemical space of viable small-molecule candidates is estimated at 10^60 — vastly larger than the number of atoms in the known universe. No enumeration approach is feasible.

This is where the concept of "Agents for Science" enters. The framework has three components:

• Prediction: Models that predict molecular structure, binding affinity, ADME properties, and toxicity at experimental-grade precision. AlphaFold 3 and its successor models fill this role.
• Exploration: Navigation of the 10^60 chemical space to find candidate molecules with desired properties. Total search is impossible; intelligent search is essential.
• Search algorithms / agents: Systems that combine predictive models and generative models — using reinforcement learning and similar approaches — to autonomously explore promising regions of chemical space and optimize molecular designs. This mirrors AlphaGo's approach: not enumerating all possible Go moves, but learning to search promising branches.

Isomorphic Labs is simultaneously building the high-resolution "map" (predictive models) and the "explorer" (generative models and search agents) that navigates it.

Data, Teams, and the "Move 37" Moment Ahead

Demis Hassabis has said that biology is not data-constrained — meaning that existing public data and experimentally accessible data already provide room for substantial AI-driven progress. Max Jaderberg extends this: in many modeling spaces, data that has existed for decades has not yet been fully exploited, because the models to extract value from it are only now being developed.

But this is not an argument against data collection. Rather, the next generation of breakthroughs may require new data generated specifically to optimize ML model training. "Machine learning-ready biological data has not really been created yet," says Jaderberg. Traditional biology experiments were not designed to maximize AI learning efficiency.

Future progress will likely require: synthetic data from physics simulation, quantum chemistry calculations, and molecular dynamics; active learning loops where models propose new data points that are then experimentally validated; and novel experimental platforms like organoid-on-chip systems that can generate high-quality in vitro data at scales that replace or complement animal testing. Isomorphic Labs does not operate its own wet lab but works with partners and contract research organizations to generate purpose-built data.

The team structure reflects the interdisciplinary challenge. No person is simultaneously a world-class drug discovery expert and a world-class machine learning expert — because the field is too new for such hybrids to have developed. Isomorphic Labs' approach is to place top experts from each domain literally next to each other, learning each other's vocabulary and building shared intuition through daily contact. Jaderberg notes that ML team members who arrive without biology or chemistry backgrounds sometimes bring fresh perspectives that cut through established dogma. "Being a little naive, intensely curious, and highly self-motivated" is the profile he values most in this environment.

The company also publishes actively, having released AlphaFold 3 model code and weights for non-commercial academic use via the AlphaFold Server — continuing the open science tradition of the AlphaFold series.

The "Move 37" Moment

What will the "GPT-3 moment" look like in AI drug discovery — the inflection point where AI output reaches or exceeds human expert quality?

Jaderberg's prediction: it will look less like GPT-3's smooth emergence and more like AlphaGo's Move 37. In the 2016 Go match that stunned the world, Move 37 was a play that professional human players initially assessed as an error. It turned out to be the decisive move — an insight that human intuition and existing frameworks simply could not generate.

In drug discovery, the equivalent moment will be when AI proposes a molecular design that human chemists would have dismissed based on intuition or convention — but which turns out to be physically and chemically correct, with superior properties. Jaderberg notes that within Isomorphic Labs' internal programs, there are already cases where AI-proposed designs have outperformed human expert judgments in experimental validation. AI as a "creative partner in scientific discovery" — not just a tool — is becoming real.

Regulatory pathways, clinical trial designs, and the broader pharmaceutical industry will need to adapt as AI-generated drug candidates increasingly enter development. Jaderberg anticipates industry convergence rather than disruption: "In five years, it will be unthinkable to do drug discovery without AI." Just as mathematics became the foundational language of science, AI will become a fundamental tool across biology and chemistry.

Summary

Isomorphic Labs' pursuit of AI drug discovery represents a genuine paradigm shift — not incremental efficiency improvement, but a fundamental rethinking of how drugs are discovered and designed.

AlphaFold 3 is a crucial milestone, but one the team explicitly frames as a beginning. The half-dozen remaining AlphaFold-level breakthroughs across molecular function prediction, ADME, and toxicity; the science agents capable of intelligently navigating a 10^60 chemical space; the interdisciplinary teams working at the intersection of ML and biology — all of these together constitute the architecture of a universal drug discovery engine.

The "Move 37" moment — when AI drug design produces an insight that human experts could not have reached on their own — is approaching. The implications for human health, and for the pharmaceutical industry, are difficult to overstate.

Reference: https://www.youtube.com/watch?v=LrMKsBtx5Bc

Related Articles

• The Reality of Working Reduced Hours After Two Maternity/Paternity Leaves – And How My View of Work Changed | TIMEWELL
• Before Taking Parental Leave (Your Own Story, Part 2): Three Must-Do Points for Taking Leave During the Busiest Season
• Committed to the Field: A First-Class Architect's Office. Finding My Own Way as the Fifth Generation of a Construction Company – Fujita Construction