Interview, Fireside Chat
The Quest to ‘Solve All Diseases’ with AI: Isomorphic Labs’ Max Jaderberg
- Isomorphic Labs was founded with the explicit ambition of building a general AI-driven drug design engine applicable across all disease areas and modalities, rather than focusing on specific targets.
- The company recently released AlphaFold 3, a breakthrough model capable of predicting the structures and interactions of proteins, small molecules, DNA, and RNA.
- Demis Hassabis's leadership in these AlphaFold advancements was recognized with the Nobel Prize in Chemistry in late 2024.
- Max Yoderberg, Chief AI Officer, previously led DeepMind research on reinforcement learning, including the breakthroughs AlphaStar and Capture the Flag.
- AlphaFold 3 utilizes a diffusion-based architecture with atomic-resolution tokenization to predict 3D coordinates for mixed molecular modalities.
- Isomorphic Labs aims to create approximately "half a dozen" AlphaFold-level breakthroughs to solve the multiple complex dimensions of drug design.
- The company estimates the potential space of drug-like molecules at 10^60, noting that even screening 1 billion molecules leaves 10^31 structures unexplored.
- To navigate this vast chemical space, Isomorphic relies on generative models and agents that can explore design possibilities without exhaustive search.
- AlphaFold 3 has shifted internal workflows from months-long lab crystallization to instant in-silico structural analysis for drug designers.
- Isomorphic currently runs internal drug discovery programs focused on immunology and oncology.
- The company has established strategic partnerships with Eli Lilly and Novartis, expanding the collaboration with Novartis after one year of success in uncovering new chemical matter.
- Isomorphic Labs does not operate its own wet labs, instead relying on partnerships and proprietary data generation to validate models.
- Max Yoderberg estimates that 60-80% of the team's machine learning scientists have no prior professional knowledge of chemistry or biology.
- The company launched the AlphaFold 3 server in November to provide open academic access to the model for non-commercial research.
- Future research directions include moving beyond static structure prediction to modeling the dynamic behavior of biomolecules in solution.
- Yoderberg envisions a "GPT-3 moment" for biology that resembles AlphaGo's "Move 37," where AI generates solutions that exhibit superhuman creativity but may be initially uninterpretable by humans.
- Predictions indicate that within five years, traditional pharma companies will be unable to design drugs without AI integration, making AI a fundamental tool for the entire industry.
- The company plans to engage with regulatory bodies to establish new frameworks for clinical trials that leverage predictive models for toxicity and efficacy.
- Yoderberg argues that while data is not currently a total bottleneck, significant opportunities exist for generating synthetic data and new "in vivo" data via technologies like organ-on-a-chip.