Derek Lowe: AI Drug Discovery Still Has Little Evidence of Clinical Impact
- A Nature Reviews Drug Discovery paper cited by Derek Lowe finds evidence of AI's clinically relevant impact remains "disappointingly limited," calling the field an absence of evidence rather than evidence that AI cannot work.
- Lowe identifies Phase II trial success rates as the outcome where AI-driven drug-development improvements should show up, since clinical trials dominate both development time and spending and Phase II has a high failure rate.
- The paper urges AI teams to shift from modeling readily available data that is unlikely to change outcomes toward problems worth solving, even when they require substantial data generation and higher upfront cost.
- Drug discovery data spans assays with many changing experimental and biological variables, and Lowe says researchers do not know how to clean and categorize it for reliable machine learning, or whether this is possible at the needed scale.
- The article distinguishes finding a ligand from producing a drug: predictions must carry from isolated proteins to cells, rodents, dog toxicology, and ultimately human clinical success.
Hacker News opinions
Drug discovery scientists already think hard about why they use each tool. AI is another tool, and drug-development timelines are longer than these methods have been genuinely effective, so measuring clinical impact will take time.
AI-made hair-loss treatments are not entirely hypothetical. I know of MINX, a slow-release oral minoxidil formulation where AI helped with formulation, while Veradermics has raised hundreds of millions for its extended-release oral product VDPHL01.
I work as a structural biologist at a mid-sized biotech and use AI every day. It makes existing work faster, like installing academic software, writing data-analysis scripts, reviewing experiment drafts, and recalling obscure formulas, but it has not produced genuinely novel ideas for me.
AlphaFold is useful as a starting model for a chimeric fusion. It turns one or two hours of searching through PDB or CIF files into a quick prompt, but that is acceleration rather than a new scientific capability.
Have you tried applying AI to more of your job in a deliberate goal-seeking way? I am curious whether you hit a real roadblock or mostly use it for the tasks where it is already comfortable.
The lack of comparable data and testability is the real problem here. I wonder whether people would share more health data if a nonprofit collected, anonymized, and governed it.
I built crohns.ai because I am not positioned to discover a drug. An AI-native clinical-trial manager evolved into a codified care protocol that helped me avoid a Crohn's flare after I lost my job, insurance, and access to Skyrizi.