Interview
AI at the Intersection of Bio | Vijay Pande, Surya Ganguli & Bowen Liu
- Computational chemistry and drug design are expected to transition from experimental validation reliance to rapid computational screening, with a specific operational shift where calculated compounds are subsequently purchased and tested by collaborators.
- The industry consensus has resolved the question of AI's viability in drug design, shifting the primary focus to implementation strategies and the physical bottlenecks of synthesizing and testing generated molecules in a lab setting.
- Deep learning is predicted to revolutionize the field by combining large datasets with self-supervised learning on unlabeled data, as exemplified by GPT-4's training on approximately five trillion unique tokens versus ESM3's approximately one trillion tokens derived from 2.8 billion amino acid sequences.
- Future medicine and materials will likely consist of novel molecules previously unseen, as current commercial offerings represent only a small fraction of the total chemical space.
- A fundamental operational shift in success rates is anticipated if AI-generated molecules achieve a screening success rate of four or five active compounds out of five screened, contrasting sharply with the current 20-30% false positive rate.
- Physics-based models are projected to outperform machine learning models specifically when test data distributions differ significantly from training data, whereas the most effective models for specific tasks are those containing the exact test example or similar data for interpolation.
- AI applications in clinical development will likely involve selecting optimal patient populations by analyzing EMR records and biomarkers to reduce heterogeneity, potentially raising clinical trial success rates from the current 20% to 30% or higher.
- Slight improvements in failure rates could counteract the "inverse Moore's law" trend where the number of successful drugs per dollar spent has historically declined.
- Personalized medicine will evolve through the use of induced pluripotent stem cells (iPSC) to generate patient-specific tissues for testing drug cardiotoxicity across diverse populations.
- A "mega foundation model" integrating biology, target understanding, trial optimization, and personalized medicine is expected to emerge, potentially serving as an "AI biologist" and "AI doctor" that embeds humans in a latent space to predict drug effects based on behavior.
- While combining data modalities like proteomics, gene expression, and wearables via a single foundation model is viewed as a seductive goal for additional insights, it remains a future objective.
- Validating generated scientific ideas is expected to remain more difficult than generating them, particularly in biology compared to computer vision or natural language processing, due to biological systems' low-dimensional structures shaped by billions of years of evolution.
- It is estimated that approximately ten years are required before AI-driven predictive models and digital humans are regularly deployed into clinical practice, with the field expected to explore previously unimaginable scientific discoveries within that timeframe.