Alphabet's Isomorphic Labs: Turning Cancer Into a Chronic, But Livable Disease

Key Concepts

• Drug Design Engine: An AI-powered system for generating new molecule designs for various diseases and modalities.
• AI Models: Predictive and generative AI capabilities used to understand protein structures, interactions, and drug properties.
• Protein Folding Problem: Determining the 3D structure of proteins, crucial for understanding their function.
• Binding Affinity: The strength of interaction between molecules, important for drug efficacy.
• In Silico Drug Design: Designing and testing drugs using computer models and simulations.
• Generalizability: The ability of AI models to be applied to novel targets and disease areas beyond their initial training data.
• Molecular Space: The vast universe of possible molecules, estimated at 10 to the power of 60.
• Reinforcement Learning: A machine learning technique used to optimize molecule design based on desired criteria.
• Human Proteome: The entire set of proteins expressed by the human genome.

Isomorphic Labs' Drug Design Engine

Isomorphic Labs is developing a "drug design engine," a machine powered by AI models to create new molecule designs for various diseases and modalities. This engine relies on multiple AI breakthroughs, not just one like AlphaFold.

Key Components:

• Structure Prediction Models: Like AlphaFold, these models predict the 3D structure of proteins and their interactions with other biomolecules (DNA, RNA, small molecules). Understanding these structures is crucial because changing the 3D structure of a system can modulate its function and alter the state of a disease.
• Binding Affinity Models: These models predict how strongly molecules bind together, which is essential for drug efficacy.
• Drug Property Models: These models assess whether a molecule is a "good drug" by predicting its safety, absorption, distribution, metabolism, and excretion (ADME) properties. This includes factors like gut absorption and cell wall penetration.

Traditional vs. AI-Driven Drug Discovery

Traditional drug design is a slow, iterative process involving:

1. Making slight tweaks to a drug molecule.
2. Synthesizing the molecule in a lab (weeks to months).
3. Testing it in a biological system.
4. Iterating based on the data.

Isomorphic Labs' approach aims to "lift that all out of the real world" by performing these steps virtually on a computer. This involves:

1. Designing molecules and testing them through AI models with near-experimental accuracy.
2. Performing multiple iterations on the computer.
3. Taking the best candidates into the lab, skipping steps and accelerating the process.

This approach aims to significantly reduce the time frame for drug discovery.

Challenges and Breakthroughs Beyond AlphaFold

While AlphaFold has provided a better understanding of the biomolecular world, significant challenges remain:

• Improving Accuracy: Further increasing the accuracy of structure prediction models.
• Understanding Interactions: Accurately predicting how molecules interact with different parts of the body, including off-target interactions that can cause toxicity.
• Navigating Molecular Space: Efficiently searching the vast molecular space (10^60) to identify promising drug candidates. Brute-force approaches are insufficient.
• Generative Models: Developing generative models that can create novel molecule designs and effectively search the entire molecular space, narrowing it down to a manageable number of candidates for lab testing.

Bridging the In Silico and Real Worlds

A key challenge is ensuring that molecules designed in silico (on the computer) work effectively in the real world (in the lab).

• Model Enrichment: Recognizing that models may not be perfect and using them to enrich the design process.
• Strong Hypotheses: Developing strong hypotheses for why specific molecular changes are being made.
• Experimental Validation: Synthesizing and testing promising candidates in the lab to validate the models' predictions.

Successes include generating novel, potent molecules with desirable properties, while failures highlight areas for model improvement.

Data Strategy

High-quality data is crucial for training effective machine learning models ("garbage in, garbage out"). Isomorphic Labs employs a comprehensive data strategy:

• Multiple Data Sources: Utilizing publicly available data and historical data.
• Bias Mitigation: Addressing biases in historical data collected for purposes other than machine learning.
• Data Generation: Creating new wet lab data specifically for training machine learning models.
• Data Quality: Investing in understanding, ingesting, cleaning, and extracting signal from data.

Generalizability

Isomorphic Labs prioritizes building generalizable technology that can be applied to any target, disease area, or modality.

• Reusable Engine: The goal is to create a drug design engine that can be used repeatedly on different targets and diseases.
• Novel Discoveries: The models should be able to discover chemical matter that has never been seen before.
• Foundation Models: Building large foundation models, like AlphaFold3, that generalize strongly to unseen systems.

Building generalizable models is more ambitious and challenging than creating targeted models for specific problems.

Focus Areas: Immunology and Oncology

Isomorphic Labs focuses its internal pipeline on immunology and oncology for several reasons:

• Tractable Clinical Trials: Clinical trials in these areas are more manageable and can be run in a shorter time frame.
• High Impact: These diseases have a significant impact on global health.
• Good Preclinical Models: These areas have good preclinical models with strong translatability to the clinic.

Outlook on Cancer Treatment

While "curing cancer" is a massive statement (cancer is a collection of diseases), the goal is to transform cancer into a chronic disease where patients can live a normal lifespan with ongoing treatment. This is seen as a stepwise process achievable in years rather than decades.

Power of the Models

The models can analyze entire families of proteins (the whole human proteome) in parallel. This is a significant departure from traditional experimental approaches, where studying proteins is time-consuming and limited in scale.

Reinforcement Learning in Drug Design

Reinforcement learning is used to optimize molecule design based on desired criteria. It involves:

1. Using generative models to create molecules.
2. Using scoring systems to evaluate the molecules.
3. Using reinforcement learning to train agents that can iteratively improve the molecules based on the design criteria.

Partnerships and Milestones

Isomorphic Labs has collaborative partnerships with Novartis and Eli Lilly.

• Challenging Targets: Novartis identified targets that had been worked on for many years with limited success.
• Dark Protein Matter: Some targets represented "dark protein matter" with no known binding molecules.
• Progress: Isomorphic Labs has made good progress on these challenging targets, identifying some of the first chemical matter binding to previously intractable proteins.

Future of Medicine

In the future, medicine could involve:

• AI-Powered Diagnosis: Using AI tools to help diagnose diseases.
• Personalized Medicine: Developing medicines tailored to individual patients.
• Disease Management: Transforming diseases into manageable conditions with ongoing treatment.

Conclusion

Isomorphic Labs is leveraging AI and machine learning to revolutionize drug discovery. Their focus on building a generalizable drug design engine, combined with a strong data strategy and collaborative partnerships, positions them to make significant advancements in treating various diseases and improving human health. While challenges remain, the potential impact of their work is substantial, with the long-term goal of solving all diseases.