AI Mines Petabytes to Speed Drug Discovery
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- Drug discovery's failure rate drives the AI push — Skolnick estimates roughly 1 in 5 drugs produce side effects that outweigh benefits, and about half of those that pass safety screens simply don't work, making AI's ability to "look at basically all available knowledge" especially valuable for flagging risks early.
- Deep learning networks like AlphaFold use transformers with billions to trillions of parameters and an "attention" mechanism to predict three-dimensional protein structures and recommend small molecules that bind to target proteins, per Skolnick.
- AlphaFold was trained on millions of protein sequences and several hundred thousand structures, enabling it to suggest molecular designs that induce specific structural shifts in disease-related proteins.
- Disease interrelationship mapping is a quietly radical AI application — Skolnick notes AI can identify co-occurring conditions (e.g., hyperthyroidism frequently leading to Alzheimer's) and trace disease trajectories back to root causes, opening paths to broad-spectrum treatments rather than narrow, one-disease-at-a-time drugs.
- Drug modalities range from small molecules like aspirin to large protein-based antibodies and gene therapies, each requiring distinct computational approaches because their scales and interaction rules differ widely, per Brown.
- AI builds on a 15-year foundation of smaller machine-learning algorithms in drug design; Brown emphasizes that AlphaFold's protein-structure predictions have expanded the toolkit rather than replaced prior computational approaches like molecular dynamics simulations, which won a Nobel Prize in 2013.
- Skolnick cautions against overpromising, stating that AI will help and accelerate discovery but is "not a substitute yet for real experiments, real clinical validation and trials."
Why it matters: AI's ability to trace disease trajectories back to root causes could shift drug development from narrow, single-disease treatments toward broad-spectrum therapies for patients who have exhausted standard options. The billion-parameter models highlighted by Skolnick function as accelerators for an industry where roughly 80% of candidates fail — but as both experts stress, clinical validation remains the bottleneck no neural network can bypass.
Ask SkimNews
