Conference Presentation, Lecture
AI is Industrializing Discovery
- The industrial revolution driven by artificial intelligence and machine learning is projected to evolve through year-over-year improvements of 10% to 20%, with a timeline to full completion estimated at two decades or longer.
- Technologies will develop to address complex biological challenges involving very small data sets, utilizing deep learning on DNA sequences to detect cancer from circulating tumor cell DNA.
- Graph convolutional neural networks are predicted to yield a huge impact in small molecule prediction, dramatically outperforming traditional random forest methods.
- One-shot learning algorithms will enable drug design to achieve high accuracy using minimal data points, shifting from unpredictable results to reliable outcomes with just one positive and one negative example.
- Pre-training schemes on unlabeled chemical data will allow algorithms to predict small molecule properties with higher accuracy while requiring significantly less data.
- Combining physics with machine learning via electron densities and conformal Schrödinger equations is expected to produce considerably more accurate results than relying on either approach independently.
- The discovery process will transition from an apprenticeship or guild model to a scaled system where humans leverage "superpowers" alongside AI and robotic data generation.
- Machine learning will predict clinical trial phases from early preclinical and phase one data, aiming to improve prediction accuracy or speed by 5% to 10%, which is valued as highly profitable.
- Modest increases in the accuracy or prioritization of clinical trial predictions will enable drugs to reach the market significantly faster and cheaper.
- Animal models and ex vivo organoid-like models will be redefined as machine learning features to enable higher accuracy in predicting human outcomes from mice without initial human testing.
- The cost of protein therapeutics, representing seven of the top ten drugs, could decrease by half, a fifth, or even a tenth through machine learning applications.
- The future biopharmaceutical sector will industrialize discovery by connecting machine learning skills with deep domain experience to scale previously bespoke human processes.