New Test Shows AI Can Help Design Drugs, But Still Makes Mistakes
AI is getting better at drug discovery, but it still can't replace human chemists.
Scientists have created a new test called MedChemBench to see if AI can help discover new medicines. The test gives AI problems from real drug research, like figuring out a molecule's structure and predicting how well it might work as a drug. The best AI model scored 84.5%, but it still made mistakes, especially when predicting how a drug would behave in the body. This is important because drug discovery is slow and expensive, and AI could help speed it up—but only if we know its limits.
In one example, AI had to predict how three similar compounds would block a cancer-related protein, cross cell barriers, and break down in the liver. The AI had to balance these competing properties, which is a core challenge in drug design. The test also checked if AI could correctly draw a molecule's structure from a description. One AI model got it wrong, drawing a ring with 15 atoms instead of 18. These kinds of errors show that AI isn't ready to work without human supervision.
The test uses 99 problems from 117 research papers and includes 761 molecular structures. It grades AI on four things: getting the structure right, explaining the science, ranking compounds, and predicting properties. While the top AI did well overall, it was weaker at predicting numerical properties and choosing the best compounds. This tells us that AI can assist with some tasks but still needs chemists to guide it. The goal is to build AI that can reliably support scientists, not replace them.
For the general public, this means AI could lead to new medicines faster, but we shouldn't expect AI to discover drugs on its own anytime soon. The technology is promising but imperfect, and human experts are still essential. As AI improves, it might help scientists focus on the most promising candidates, saving time and money. But for now, it's a tool to aid research, not a replacement for human judgment.
- A new test called MedChemBench measures how well AI can solve real drug-discovery problems.
- The best AI scored 84.5%, but struggled with predicting how drugs behave in the body.
- AI still makes mistakes, like drawing a molecule incorrectly, so human oversight is crucial.
Why It Matters
AI could speed up drug discovery, but its errors mean we still need human experts to ensure safety and effectiveness.