Why is QSAR so far behind other forms of machine learning, and what can be done to close the gap?
QSAR models struggle with extrapolation compared to conventional ML tasks like image recognition. Machine learning generalizes effectively when structured to align with its problem domain, suggesting that improving QSAR models may close this gap in drug discovery.
100 AI-generated molecules are worth a 1,000,000 molecule high-throughput screen
Generative AI in drug discovery is showing promise by optimizing molecule searches beyond traditional methods. Variational AI’s Enki algorithm created 100 AI-generated molecules that outperform 1,000,000 in conventional high-throughput screening, highlighting AI’s efficiency in exploring chemical space.
Applicability domains are common in QSAR but irrelevant for conventional ML tasks
Traditional QSAR models are limited to interpolation within known chemical spaces, restricting drug discovery. In contrast, modern machine learning excels at extrapolation, opening new possibilities for exploring untapped chemical compounds and enhancing hit discovery.