AI Filters Out False Leads in the Search for Tuberculosis Drugs

A team at Texas A&M AgriLife Research led by James Sacchettini has built an artificial intelligence model aimed at solving one of the most time-consuming problems in tuberculosis drug discovery: telling compounds that are genuinely effective apart from ones that merely look promising in the lab.

Tuberculosis, caused by the bacterium Mycobacterium tuberculosis, is among the oldest diseases known to medicine, yet it still evades many therapies. The TB bacterium has a thick, waxy coat that blocks most drugs from getting inside the cell, and it grows exceptionally slowly, meaning a single lab experiment can take months instead of weeks.

The false lead problem

Finding a new drug starts with screening thousands of chemical compounds for activity against the bacterium. Many candidates that look promising in the first test turn out to be useless under closer scrutiny. Researchers call these "nuisance molecules": they clump together in solution, interfere with the assay's chemical signal, react instead of binding to the target, or latch onto many different structures at once instead of acting selectively.

Weeding out these false leads by hand consumes months of a research team's time before testing of genuine drug candidates even begins. Sacchettini described the problem bluntly, noting the enormous time cost of such missteps.

These nuisance molecules cost us a ton of time - James Sacchettini, Texas A&M AgriLife Research

How CAGE-Fusion works

CAGE-Fusion is a model trained on published chemical screening data that classifies compounds by four types of problematic behavior: molecular clumping, assay signal interference, excessive chemical reactivity, and binding to multiple targets instead of one. In benchmark tests where the model had to distinguish one suspect compound from one clean one, accuracy reached about 94 percent, though performance varies by category. Highly reactive compounds are the easiest to catch, while those binding multiple targets at once prove the hardest.

The team stresses that the tool isn't meant to replace chemists or point to a finished drug, but to screen out dead ends before a lab invests weeks of work in them.

We're not counting on AI to give us exactly the right answer. But it can tell us what's not worth working on, which in turn suggests what we should focus on - James Sacchettini, Texas A&M AgriLife Research

A knowledge base instead of scattered data

Alongside CAGE-Fusion, Sacchettini's lab is developing DAIKON, an open platform launched in 2023 that tracks the research history of a given molecular target, from gene discovery through years of chemistry work, all in one place. The entire Tuberculosis Drug Accelerator, a consortium of labs and companies backed by the Gates Foundation, relies on the platform. Siddhant Rath and Saswati Panda, from the same team, built an additional chat-based tool that lets researchers pull up prior results and materials on a given molecule within seconds, instead of searching archives by hand.

Why it matters

Tuberculosis kills more people each year than any other infectious disease, and it hits low-income regions hardest, where access to new therapies is often limited. Drug-resistant cases and HIV co-infections require even longer treatment, straining healthcare systems further. Automating the process of screening out false drug candidates doesn't guarantee a therapeutic breakthrough, but it meaningfully cuts the time research teams lose chasing dead ends, which matters especially for a slow-growing bacterium like this one.

For readers following AI's expansion into science, this is another example of machine learning models moving into preclinical research stages that once required purely painstaking lab work. Similar approaches, combining machine learning with classic chemical screening, are increasingly showing up at research centers working on new antibiotics as well.