Researchers at Texas A&M AgriLife have built artificial intelligence tools that help scientists quickly discard false-positive drug candidates during tuberculosis drug discovery, cutting months of wasted effort in a field where experiments are notoriously slow. The tools-a model that identifies nuisance molecules in screening assays and a search system that surfaces knowledge across a global research consortium-address a persistent bottleneck in the fight against the world's deadliest infectious disease.
According to the World Health Organization, tuberculosis kills more people each year than any other infection. The bacterium's thick, waxy outer coating blocks most drugs, and its slow growth means a single experiment can take months, compared to days with common bacteria. "To do a tuberculosis experiment takes months sometimes, where it can take a week with staph or strep," said James Sacchettini, Ph.D., the Rodger J. Wolfe-Welch Foundation Chair in Science at Texas A&M AgriLife Research. "That's part of the reason why drug discovery pipeline has been relatively slow. It's a perfect area to work with AI."
Sacchettini's lab is part of a broader movement in biomedical research to apply AI for Science & Research to early-stage drug screening, where models can now flag compounds that look promising but lead nowhere.
Building a shared data foundation
Before the AI models could work, the lab first tackled a more basic problem: scattered data. Academic labs often store results across network drives, slide decks, and individual memory. The team built DAIKON, an open-source platform published in 2023, that tracks a drug target from the gene level through years of chemistry work in one place. The Gates Foundation-supported Tuberculosis Drug Accelerator (TBDA) consortium adopted it across its entire partnership of labs and companies. Both of the new AI systems plug directly into DAIKON.
Filtering out false leads with CAGE-Fusion
Early in development, high-throughput screens generate thousands of candidate compounds, many of which interfere with the test itself rather than genuinely hitting the target. "These 'nuisance molecules' cost us so much time," Sacchettini said. To solve this, Sacchettini, AgriLife Research scientist Siddhant Rath, and colleagues developed CAGE-Fusion, an AI model that learns from published screening data to sort compounds into four trouble categories: aggregators that clump together, compounds that deceive the assay's chemical signal, reactive molecules, and those that stick to many targets instead of one.
In head-to-head comparisons of one nuisance compound and one clean hit, the model ranks the nuisance compound as more suspicious roughly 94% of the time. It is most accurate at catching reactive compounds and less adept at identifying promiscuous binders. Inside DAIKON, CAGE-Fusion runs automatically on incoming data, flagging likely problems before a compound advances to costly later stages. "The model can walk you through the process and show you which regions of the molecules it thought were problematic," Rath said. For biochemists designing screening assays, understanding how models like CAGE-Fusion categorize nuisance compounds can prevent months of wasted bench work-a skillset covered in a dedicated AI Learning Path for Biochemists.
Making years of consortium data searchable
Rath and colleague Saswati Panda also created an AI system that brings order to the TBDA consortium's accumulated data. The partnership's years of documentation span thousands of molecular structures, but until now there was no simple way to trace a molecule's history across projects. The new system indexes TBDA data along the entire drug discovery pipeline. Researchers can query it through a chat interface. "It will tell me in a matter of seconds who presented something, what they said about it, and give me the presentation so I can see the slides," Sacchettini said.
Why this matters for Science and Research
The tools demonstrate a practical AI strategy for drug discovery-one that prioritizes identifying dead ends over trying to predict the perfect drug. Sacchettini summed up the philosophy: "We're not hoping for AI to give us the exact right answer. But it can tell us what not to work on, which then informs us what we should be working on. And it really is a big time saver."
For research scientists working on neglected diseases or any target with slow assay cycles, the same approach is replicable. Build a centralized, open-source data platform, layer on machine learning models that catch common screening artifacts, and add natural language tools that let teams query institutional knowledge directly. As these models from Texas A&M move from prototype to daily use inside a major consortium, they show how AI can cut months of dead-end work without requiring a perfect prediction.
Your membership also unlocks:
