In a direct comparison, a medicinal chemist was better than an AI model at evaluating the synthesizability of 30,000 compounds. The chemist's intuitive, "liability-spotting" approach highlights the continued value of expert human judgment and the need for human-in-the-loop AI systems.
The groundbreaking AI-driven discovery of antibiotics is relatively unknown even within the AI community. This suggests a collective blind spot where the pursuit of AGI overshadows simpler, safer, and more immediate AI applications that can solve massive global problems today.
Developing an antibiotic is costly, but its use is short-term and new drugs are held in reserve, making them unprofitable. This market failure, not a lack of scientific capability, has caused pharmaceutical companies to exit the space, creating a worsening global health crisis.
MIT Professor Jim Collins estimates a $20 billion investment could fund the R&D and clinical trials for 15-20 new antibiotics, solving the crisis for decades. This cost is a fraction of recent tech investments, framing an existential threat as a solvable, relatively affordable problem.
The AI-discovered antibiotic Halicin showed no evolved resistance in E. coli after 30 days. This is likely because it hits multiple protein targets simultaneously, a complex property that AI is well-suited to identify and which makes it exponentially harder for bacteria to develop resistance.
To overcome a small training set, researchers discretized continuous growth inhibition data into a binary (yes/no) classification. This simplified the learning task, enabling the model to achieve high predictive power where a more complex regression model would have failed due to insufficient data.
Professor Collins' AI models, trained only to kill a specific pathogen, unexpectedly identified compounds that were narrow-spectrum—sparing beneficial gut bacteria. This suggests the AI is implicitly learning structural features correlated with pathogen-specificity, a highly desirable but difficult-to-design property.
Models designed to predict and screen out compounds toxic to human cells have a serious dual-use problem. A malicious actor could repurpose the exact same technology to search for or design novel, highly toxic molecules for which no countermeasures exist, a risk the researchers initially overlooked.
Professor Collins’ team successfully trained a model on just 2,500 compounds to find novel antibiotics, despite AI experts dismissing the dataset as insufficient. This highlights the power of cleverly applying specialized AI on modest datasets, challenging the dominant "big data" narrative.