AI-Driven Drug Discovery Faces Data Quality and Integration Hurdles
Since the 1950s, the expenses associated with drug discovery have increased twofold every nine years, with the development of new medications requiring 10 to 15 years and costing between $1 billion and $2.5 billion, while experiencing failure rates exceeding 90%. Artificial intelligence seeks to enhance success rates and shorten development times, but it relies on comprehensive data and integration within laboratories. Paul Belcher from Cytiva emphasizes that while AI can facilitate predictive design, it cannot yet accurately forecast kinetics or developability, making lab validation essential. He highlights the issue of the 'data wall,' as many AI models trained on public datasets yield diminishing returns. Belcher references Elisabeth Bik's research on altered images in biomedical literature. No AI-generated drug has achieved full FDA approval yet, but Belcher anticipates this will occur within two to three years.
Key facts
- Drug discovery costs have doubled every nine years since the 1950s (Eroom's Law).
- Bringing a new drug to market takes 10-15 years and costs $1-2.5 billion, with failure rates over 90%.
- AI is used for hit identification, shifting from empirical screening to predictive design.
- AI cannot yet predict kinetics or developability; candidates must be validated in the lab.
- Public datasets for AI training suffer from publication bias, focusing on positive results.
- Elisabeth Bik found almost 4% of biomedical papers contained duplicated or manipulated images in 2016.
- Cytiva's Image Integrity Checker uses secure hash algorithms to detect image tampering.
- No AI-discovered drug has received full FDA approval yet; Belcher expects that in 2-3 years.
Entities
Institutions
- Cytiva
- MIT Technology Review
- Stanford
- FDA